X Bookmarks: AI Agents & Tools (Jan 2025 – Jun 2026)

Compiled AI/agent signals from Kevin's X bookmarks. Grouped by sub-theme, sorted by engagement within each group. Excludes bookmarks already processed into dedicated pages (Knowledge Brain System, Cursor Orchestrate Skill, Chris Tate, Browser Testing Skills).

Deep Review 2026-07-06

Agent participation, docs, and skill surfaces

Takeaway: The new batch is unusually coherent: agents need better participation artifacts, stronger source-bound docs, and more explicit command/tool surfaces. Staying in the Loop with Agents absorbs Geoffrey Litt's thread as "participation artifacts" for human judgment while agents write code. OpenWiki, X MCP, and Agent-Docs Mesh cover the source-retrieval lane: OpenWiki turns repos/docs into agent-readable material, while X's hosted MCP endpoints expose X search/bookmarks/docs to agents with a clear OAuth boundary. BuilderIO Skills and Matt Pocock's Skills (Skills for Real Engineers) absorb the skill-authoring cluster: visual recaps/design-system indexing on the Builder side; rename/watch-queue and review-vocabulary movement on the Pocock side. Source: X bookmark artifact audit, 2026-07-06

Runtime, browser, and model-loop signals

Takeaway: Hermes Agent (Nous Research) gets the biggest runtime update: browser extension, operator field notes, sovereignty/privacy rows, and benchmark-required web-extraction claims. Browser Testing Skills adds Browser Use CLI 3.0 as a coding-agent browser-control surface while preserving Kevin's local Browser Testing Skills / Browser Harness - Self-Healing Browser Automation default. LLM Council absorbs TRINITY as a tiny-coordinator research watch, while Harness Engineering absorbs ARC AGI harness rows and the "hard to eval" caveat. AI CLI (Vercel) and Vercel carry the Vercel terminal/platform side: model catalog, audio commands, Agent Runs in CLI/MCP, dry-run deployments, and konsistent. Source: X bookmark artifact audit, 2026-07-06

Deep Review 2026-07-03

Deep Review 2026-07-04

Durable sessions, thin brains, and marketing agents

Takeaway: OpenComputer Durable Agent Sessions now has a current SDK/source snapshot: @opencomputer/sdk@0.11.2, GitHub diggerhq/opencomputer at eccef9d, and a hosted-service page that separates the model-running brain from the sandbox hands. Knowledge Brain System and Garry Tan now preserve the "Thin Harness, Fat Skills" self-thread provenance while keeping the root "memory is markdown" tweet as the durable citation. Graphed was created from Cody Schneider's self-thread as a hosted marketing-agent workflow reference; local execution still routes through Writing and Content Skills and Kevin's SEO/GEO skills. Source: X bookmark artifact audit, 2026-07-04

Taste automation as market pressure

Takeaway: Taste as a Moat and Taste Is the Differentiator (on Shared Primitives) now carry the taste.ai funding reaction as a market warning: automating taste only differentiates if the system owns selection, rejection, constraints, and proof. A tool that merely makes generic output harder to spot turns slop into a harder-to-detect commodity. Source: X/@VitusDoesThings, 2026-06-17

Eve launch source review

Takeaway: Vercel Eve source review confirms the original @vercel launch bookmark still maps to the current public package: eve@0.19.0, Node >=24, Apache-2.0, and a Git tag at 923170b0. The launch claim remains file-native agent structure plus production harness features: durable sessions, sandboxed compute, approvals, subagents, connections/channels, tracing, and evals. Source: X/@vercel, Vercel blog, npm registry, and GitHub API, 2026-07-04

Agent framework checklist

Takeaway: Vercel Eve now carries Chris Tate's compact "agent framework" checklist as an audit lens for Eve-class tools: data, evals, tools, skills, schedules, subagents, durable workflows, sandboxed execution, connections, and channels. The durable point is not the branding; it is the missing-capability test. A framework that omits evals, durable runs, sandboxing, or explicit connection/channel boundaries is probably still a demo harness. Source: X/@ctatedev, 2026-06-17

Eve Nuxi migration signal

Takeaway: Vercel Eve gets an external adoption signal from @hugorcd migrating Nuxi, the Nuxt Agent, to Eve. The bookmark is not new feature authority - current source authority remains eve@0.19.0, tag 923170b0, HEAD e2d7c577 - but it matters because a Nuxt-adjacent agent author is testing Eve's file-native model against a real framework-agent codebase. Source: X/@hugorcd and Vercel Eve page, 2026-07-04

Eve observability and 24-hour sandboxes

Takeaway: Vercel Eve now carries Vercel's Agent Runs observability update: Eve sessions get dashboard traces with trigger, duration, token usage, model calls, tool calls, runtime-error correlation, developer/business views, retention limits, and optional OpenTelemetry export. Agent Sandboxes and Persistent Sandbox Thesis carry the separate Vercel Sandbox duration update: uninterrupted sessions can now run up to 24 hours, up from 5 hours, making long-lived agent workflows a mainstream platform target when paired with persistent sandboxes. Source: X/@vercel_dev and Vercel changelog review, 2026-07-04

Vercel Connect public beta

Takeaway: Vercel Connect now carries the source-reviewed public beta bookmark and current npm snapshot. The durable product boundary is task-scoped connector credentials: apps and agents request short-lived tokens for Slack, GitHub, Salesforce, and custom services at runtime instead of storing long-lived provider secrets. Current package authority is @vercel/connect@0.3.2, published 2026-07-02. Source: X/@vercel_dev, Vercel changelog, and npm registry, 2026-07-04

Agent definition and tailnet operator stack

Takeaway: Dedalus SDK now carries Dedalus's compact agent definition as product-positioning evidence: Agents := Model(s) + Tool(s). The reviewed image is a brand card with two interlocked rings labeled models and tools, so implementation claims still route through current SDK docs. Tailnet Agent Mesh - Reaching Your Agents from Anywhere now preserves @0xSero's stack-ranking bookmark as field evidence: Tailscale is the first layer, VibeProxy is the subscription/API proxy, and Codex App is the agent control surface. The linked setup status is still X/Jina-blocked, so no extra setup claims were promoted. Source: X/@dedaluslabs and local image review, 2026-07-04; Source: X/@0xSero and enriched artifact-reply review, 2026-07-04

VisionClaw and runaway agent boundaries

Takeaway: VisionClaw graduates as the wearable-agent capsule: Meta Ray-Ban / phone-camera context streams to Gemini Live, with optional tool calls forwarded through an OpenClaw gateway. Treat it as public research/prototype code with Meta Wearables terms and hardware/privacy caveats, not a default install. Agent Legibility now owns @kevskgs' Cursor runaway-response artifact as a control-loop failure: once "I should stop" becomes more generated output, the harness needs external budgets, truncation, and summarize/ask gates. Source: X/@ihtesham2005 and GitHub review, 2026-07-04; Source: X/@kevskgs and local media review, 2026-07-04

Claude Tag, app-builder loops, workflow durability, and Supabase Auth latency

Takeaway: This slice split four superficially similar "agent wins" into their actual owners. Claude Tag is the product page for Slack-native shared Claude work: scoped channel memory, tools/data/codebase access, ambient follow-up, and Team/Enterprise controls. Harness Engineering owns the Anthropic planner/builder/evaluator loop from the AI Engineer Europe clip and official harness-design post: the durable insight is role separation plus evaluator-negotiated proof, not just using more agents. Claude Code Dynamic Workflows and Durable Agent Execution Architecture now preserve the workflow durability boundary: dynamic workflows are deterministic JS orchestration with /resume-style subagent recovery, not a Temporal replacement. Security and Review Skills and the executable Supabase Auth skill now carry the getClaims() vs getUser() performance rule, with the security caveat that claims can still be stale for sensitive authorization. Source: X bookmark artifact audit, 2026-07-04

Council fan-out and vulnerability harnessing

Takeaway: Two diagram artifacts now have canonical homes. LLM Council owns EXM's council setup as a fan-out/fusion protocol: a router selects one model, while a council runs several role-diverse models in parallel and uses a judge/frontier model to fuse the result. Harness Engineering and Project Glasswing own Eugene Yan's Cloudflare diagram: vulnerability discovery works as a staged harness, not a generic "ask a coding agent to audit the repo" prompt. Source: X/@EXM7777 and local image review, 2026-07-04; Source: X/@eugeneyan and Cloudflare Blog review, 2026-07-04

agentOS lightweight agent runtime

Takeaway: agentOS now owns Rivet's v0.2.0 launch as the lightweight WebAssembly/in-process VM route for agent execution. The local video artifact claims faster cold starts, lower memory, lower run cost, any-agent/BYO-agent support, multiplayer workflows, improved filesystem/limits, and one-prompt deploy to Rivet Cloud. The routing boundary is explicit: agentOS is for resident command loops and controlled filesystem/process semantics; Agent Sandboxes still covers full browser/dev-server/native dependency environments. Source: X/@rivet_dev, GitHub, npm, local video review, 2026-07-04

GBrain v0.9, Hermes Bible, and Princeton TD readability

Takeaway: Three older rows now have durable routing. Knowledge Brain System owns Garry Tan's v0.9.0 artifact as the code-plus-skills design rule: skills carry judgment, deterministic commands handle mechanical work. Hermes Bible is now a dedicated page for the unofficial Hermes docs/search surface, while Hermes Agent (Nous Research) remains the official source authority. Princeton Tower Defense captures the day 7 readability artifact: tower-defense UI needs visible state, selected-unit detail, and healing/buff observability before it needs prettier chrome. Source: X/@garrytan and local image review, 2026-07-04; Source: X/@iamlukethedev and Hermes Bible review, 2026-07-04; Source: X/@kevskgs and local image review, 2026-07-04

Miles Deutscher source-only roundup triage

Takeaway: Reviewed as a source-triage row, not a standalone article. The resolved X Article body was empty in the enriched record, and the self-thread mixed three independent shares: a Claude skills automation guide, an open-weight model usage note, and an AI World Cup prediction system. Existing pages already cover the durable agent-skills and open-weight routing lanes through Skills.sh - Agent Skills Registry, Browser Testing Skills, OpenRouter, and Local Coding Agents; the prediction-system reply is not source-sufficient without the underlying build details. Source: X/@milesdeutscher and enriched record review, 2026-07-03

ZeroEntropy/GBrain and local runtime routing

Takeaway: ZeroEntropy is the source page for ZeroEntropy's current embedding/reranking stack, while Knowledge Brain System records Garry Tan's product-level routing claim: GBrain recommends ZeroEntropy for embedding/reranking over OpenAI and Voyage AI. Current source checks preserve the exact authority boundary: ZeroEntropy docs expose zembed-1, zerank-2, Python/Node SDKs, US/EU endpoints, npm 0.1.0-alpha.10, PyPI 0.1.0a11, and Hugging Face SHA snapshots; GBrain's repo snapshot remains on the existing GBrain page. Source: X/@garrytan, 2026-05-23; Source: ZeroEntropy docs/npm/PyPI/Hugging Face, 2026-07-03

Ahmad's Ollama/OpenClaw post is best kept as a routing heuristic, not a new page. If the task is local model serving, compare Ollama, llama.cpp, LM Studio, and OpenAI-compatible local servers through Local Coding Agents and receipts; if the task is a persistent self-owned runtime, Hermes Harness (Nous Research) is the stronger current capsule, while OpenClaw Harness remains an alternate gateway/runtime pattern. Current primary-source check: llama.cpp release b9870 was published 2026-07-03, Ollama latest release is v0.31.1, and Hermes remains v0.18.0 / v2026.7.1. Source: X/@TheAhmadOsman, 2026-06-22; Source: GitHub ggml-org/llama.cpp, ollama/ollama, NousResearch/hermes-agent, 2026-07-03

GStack Confusion Protocol

Takeaway: Quality Review Stack now carries Garry Tan's Confusion Protocol artifact as ambiguity-gate evidence. The reviewed image says workflow skills should stop and ask when they hit an ambiguous decision instead of guessing; it also names Hermes host support, GBrain-aware skills, slop:diff, skill triggers, CEO-review reinforcement, and investigation history. Current source authority still comes from the GitHub repo and changelog, not the screenshot's historical v0.18.0 label. Source: X/@garrytan and local image review, 2026-07-03

Origami GTM agent and content skill graph

Takeaway: Origami now has artifact-level source authority: the reviewed video shows a GTM agent workspace with replies, sent-count progress, warm-intro context, and a recommendation to re-engage 5,497 free users with a draft campaign; current docs expose Data API v1 plus Agent API v2 beta with async runs and origami-lite/origami-max model selection. Social Content OS now carries @DeRonin_'s content-production skill graph as architecture evidence: content automation should be a graph of markdown instruction nodes for platforms, audience profiles, voice, hooks, workflows, cadence, and metrics, not one giant "write content" prompt. Source: X/@origamichat and local video review, 2026-07-03; Source: Origami docs, 2026-07-03; Source: X/@DeRonin_ and local image review, 2026-07-03

Raindrop Triage as MCP agent

Takeaway: MCP as the Integration Standard now carries Raindrop Triage as the agent-as-tool pattern. The durable design decision is to expose a task-specific investigator through MCP when the agent harness itself is valuable: production-data reasoning, personalization, and context handoff from Slack/web into coding agents. This is not a generic "wrap every API in an agent" recommendation; use it when the wrapper materially improves accuracy, triage speed, or cross-agent handoff. Source: X/@benhylak, 2026-05-05; Source: Raindrop Triage blog, 2026-05-04

Agent debugging and human judgment

Takeaway: Agent Legibility now carries Rauch's root debugging thread instead of only the human-judgment follow-up. The root cause pair is useful: agents are hard to debug because model outputs are intentionally nondeterministic and the surrounding system is a multi-step distributed workflow. The follow-up moves the human role upstream: choose what to build, choose architecture, decide whether to regenerate or reuse, and manage technical debt. Source: X/@rauchg, 2026-06-27; Source: X/@rauchg, 2026-06-27

Crustdata, OpenRouter MCP, and static ad workflow

Takeaway: Three parked rows now have focused homes. Crustdata is the B2B fresh-data primitive for agentic GTM workflows: company/person/job/social data through API and MCP, plus the Slack inbound-enrichment pattern from the self-thread. OpenRouter now carries the OpenRouter MCP server as a coding-agent development assistant for live model/pricing/docs lookup and test inference, while LLM Council captures the screenshot's transferable protocol: small role-diverse council, final editor, not majority-vote sprawl. Higgsfield MCP absorbs the Nano Banana static-ad examples as composition evidence for Claude plus Higgsfield-style creative workflows, not as model/version authority. Source: X/@chrispisarski, 2026-06-18; Source: X/@lennox_saint, 2026-06-26; Source: X/@alexgoughcooper and local image review, 2026-07-03

Open Higgsfield AI source review

Takeaway: Open Higgsfield AI stays a skeptical self-hosted creative-UI reference, not a verified Higgsfield replacement. The viral tweet claimed 200+ models, the GitHub repo description currently markets 20+ models, but the README still says image generation is paused until provider credits are added. Current source snapshot: ClabstreamTeam/Open-Higgsfield-AI main at 44788b5, no Git tags, 76 stars, 20 forks, and no license detected by GitHub API. Source: X/@ihtesham2005 and GitHub review, 2026-07-04

Deep Review 2026-07-01

AgentMail

Takeaway: AgentMail graduates from coverage-floor promo to agent communications infrastructure. The durable capability is real two-way inboxes for agents, with API, SDK, CLI, hosted MCP, docs-as-Markdown, webhooks/WebSockets, threads, drafts, labels, custom domains, deliverability, and idempotency. Startup Program pricing is useful context, but routing should be based on the inbox/API/MCP surface and side-effect policy, not the promo. Current source snapshot: agentmail@0.5.13, agentmail-cli@0.7.12, agentmail-mcp@0.2.2; hosted MCP server https://mcp.agentmail.to/mcp; Free/Developer/Startup tiers at 3/10/150 inboxes. Source: X/@agentmail, 2026-06-16; Source: AgentMail docs/pricing/npm/GitHub, 2026-07-01

Deep Review 2026-07-02

Higgsfield agency workflow and Hermes plugin guide

Takeaway: Higgsfield MCP now carries the root 18-minute "marketing agency in your laptop" artifact: Claude coordinates research, production planning, UGC/carousel/video creative, approval, and Meta Ads handoff. Route it as hosted creative-agent workflow evidence, not a local skill import. Hermes Agent (Nous Research) now carries the Miles Deutscher custom-plugin guide and official docs: use Hermes skills for instructions and orchestration over existing tools; use plugins when the runtime needs custom tools, hooks, integrations, API keys, binary data, streaming, or project/team-specific behavior. Current source authority: Higgsfield CLI @higgsfield/cli@1.1.1; Hermes Agent v0.18.0 / v2026.7.1, main HEAD 30e947e. Source: X/@higgsfield_ai, 2026-05-18; Source: X/@milesdeutscher, 2026-06-07; Source: Higgsfield/Hermes docs and local video reviews, 2026-07-02

GStack UX foundations and Vercel voice agents

Takeaway: Two rows now have focused homes. Quality Review Stack gets the UX capability artifact: GStack's design skills inject Steve Krug-style scanning, satisficing, goodwill, navigation, and trunk-test guidance, then run six usability tests, first-person narration, $B ux-audit extraction, heatmap screenshots, and token ceiling checks. No new local skill import is needed because gstack-design-review is already the maintained route. Vercel AI SDK gets the AI Gateway voice-agent lane: realtime transport, speech generation, transcription, server-minted browser tokens, provider routing, observability, spend controls, and BYOK. Source: X/@garrytan and local artifact review, 2026-07-02; Source: X/@vercel, 2026-06-29; Source: Vercel blog, 2026-06-29

GBrain Minions

Takeaway: Knowledge Brain System v0.11 Minions is durable-agent-execution evidence: long-running sub-agent tasks become Postgres/PGLite jobs with progress streams, transcripts, pause/resume/steer, worker supervision, and gbrain jobs list observability. The reviewed artifact claims the tested path moved from sessions_spawn gateway timeout to 753ms job dispatch, zero token cost, 100% success, and about 2MB memory/job. Route the transferable pattern to Durable Agent Execution Architecture and Agentic Infrastructure: queue the work, make it idempotent, limit concurrency, and keep operator-visible job state. Source: X/@garrytan and local artifact review, 2026-07-02

Giga Scout, Browser Use Box, and YC AI Stack

Takeaway: Three rows now have durable tool/resource homes instead of hub-only notes. Giga Scout is the KPI-directed agent-company product proof: define a business metric, learn from customer conversations, file fixes as policy/tooling/knowledge, test safe changes on slices, and escalate risky changes to humans. Browser Use Box (Bux) is the source-owned always-on browser-agent box: Claude Code, Chromium/browser-harness, Browser Use Cloud, Telegram, a web terminal, and persistent browser state on a machine the user owns. YC AI Stack is a student-founder credit bundle, useful as prototype budget for AI developer tools but not as an architecture dependency because eligibility and vendor credits can change. Source: X/@GigaAI, 2026-06-29; Source: GitHub browser-use/bux, 2026-07-02; Source: Y Combinator, 2026-01-16

Hermes LLM-Wiki, Unreal MCP, and ml-intern

Takeaway: Three older agent/tool rows now have artifact evidence and current source authority. Hermes Agent (Nous Research) packages Karpathy's LLM-Wiki pattern as a bundled research skill (skills/research/llm-wiki, version 2.1.0): the reviewed graph image shows a Nous research vault clustered around GRPO/RL post-training topics, and the live repo is now v0.18.0 / v2026.7.1 at 8b1ad38ecb8e5002f74bd503efabd1009c84ff56. MCP as the Integration Standard now records Unreal MCP as an experimental creative-tooling example, including Epic's local/no-auth/game-thread-serialized caveats. ML Intern - Automated ML Research Agent is pinned to untagged main at 550a209701701e6a9ac7cac70b8dbd508822d467, project version 0.1.0, Apache-2.0, with HF Inference Providers, LiteLLM local model routing, and optional HF Space sandbox tools. Source: X/@Teknium, 2026-04-07; Source: Hermes docs/GitHub, 2026-07-02; Source: X/@UnrealEngine, 2026-06-17; Source: Epic Unreal MCP docs, 2026-07-02; Source: X/@akseljoonas, 2026-04-21; Source: GitHub huggingface/ml-intern, 2026-07-02

Lightpanda and OpenRouter Fusion source refresh

Takeaway: Two older high-signal agent/tool rows now have current source authority and artifact observations. Lightpanda remains the fast, low-memory browser engine for cooperative agent browsing: the reviewed screenshot showed the launch-era 18K-star README and 11x/9x benchmark framing, while the live repo is now 0.3.4, HEAD 60aacfa2b064877c3eecddf98a5dd61048c1bdd8, AGPL-3.0, 31.5K stars, and 1.4K forks. OpenRouter Fusion remains a panel-and-judge route, not a new base model: the reviewed chart preserves the DRACO benchmark ordering, while live docs/model pages still expose the same three entry points, openrouter/fusion, openrouter:fusion, and the fusion plugin. Source: X/@ihtesham2005, 2026-03-15; Source: GitHub lightpanda-io/browser, 2026-07-02; Source: X/@OpenRouter, 2026-06-13; Source: OpenRouter docs/model page, 2026-07-02

Taste imagegen and Anthropic marketing video artifacts

Takeaway: Taste Skill Framework (Leonxlnx/taste-skill) gets the imagegen workflow evidence from Lex Lin's "Forma" demo: generate one horizontal image per section, use the reference board before code, and treat the upstream imagegen-frontend-web source as a candidate/reference only because Brin currently returns caution for both repo and skill routes. Austin Lau gets the Anthropic marketing interview artifact as person/workflow evidence: one growth lead used Claude Code over ad CSVs, replacement generation, email, SEO, paid search, and paid social rather than a new marketing tool route. Source: X/@LexnLin, 2026-05-02; Source: GitHub Leonxlnx/taste-skill, checked 2026-07-02; Source: X/@cgtwts, 2026-03-14

Deep Review 2026-06-30

Harness curriculum, source-leak provenance, and portable agent sessions

Takeaway: Learn Harness Engineering is still the canonical outside curriculum for Kevin's harness stack: course repo 2fdc46b20a3da696e1657a34fadfb5fd9e8acbb1, awesome list f84f1701974cf1ad67dd774b025b33e613275cee, and both repos are reference/course material rather than executable dependencies. Claude Code Source Leak now owns the iamfakeguru employee-gate artifact: keep the verification lessons, but do not import a claimed CLAUDE.md bypass because USER_TYPE === 'ant' is build-time/dead-code-eliminated in distributed builds. Omnara graduates from coverage floor to a focused tool page as evidence for portable agent sessions: the durable unit is the dirty worktree plus conversation/runtime state, not a nicer VPS. Source: X/@_vmlops, 2026-05-17; Source: X/@iamfakeguru, 2026-03-31; Source: X/@ishaansehgal, 2026-05-13

Fable availability and Browser Use QA

Takeaway: The Fable launch bookmark stays important as a capability and launch-artifact reference, but Claude Fable 5 now records Anthropic's official 2026-06-12 suspension notice. Do not route future work to Fable/Mythos without checking current access. The Browser Use QA bookmark graduates from raw research into Browser Testing Skills and Browser Automation and QA: it is an evaluator pattern for pairing a text-only builder, such as GLM 5.2, with multimodal QA subagents that review a rendered website, find bugs, judge aesthetics, and return scored evidence. Source: Anthropic, 2026-06-12; Source: X/@claudeai, 2026-06-09; Source: X/@browser_use, 2026-06-20

Vercel Eve and Zero language

Takeaway: Vercel Eve and Zero (Programming Language) now have deep-review decisions. Eve's useful structure is the file-native agent graph: instructions, tools, skills, subagents, connections, schedules, evals, package metadata, and traces on disk. The critique bookmark is retained because it catches the failure mode: a folder of markdown is not a framework unless the harness assigns runtime semantics, approvals, scoped delegation, observability, and CI evals. At the 2026-07-01 review, Eve source authority was eve@0.18.0, Apache-2.0, main HEAD/tag 67112781d3700b3024bfa9b0013eee1253814780; the current Eve snapshot is maintained in Vercel Eve. Zero is promoted to a tier-2 design reference, not a production language dependency: graph patching, stable handles, typed repairs, JSON diagnostics, and explicit World capabilities are the reusable agent-native language ideas. Current Zero source authority is GitHub release v0.3.4, Apache-2.0, main HEAD afcc72da649fe4d4c670ac1489c2197d37436051; there is no correct npm package under zerolang, @vercel-labs/zerolang, or @vercel/zerolang. Source: X/@IceSolst, 2026-06-18; Source: X/@ctatedev, 2026-05-15; Source: GitHub/npm, 2026-07-01

agentcn

Takeaway: agentcn graduates from coverage-floor row to focused page. It is a shadcn-compatible registry for Eve/Flue agent recipes, useful as a reference for source-owned agent workflows: deep search, PR review, Slack agents, browser agents, PDF chat, docs chat, database chat, and company knowledge. Import recipes only when the target app wants Eve/Flue; otherwise use it as a recipe-boundary reference. Source: X/@alaymanguy, 2026-06-20; Source: agentcn docs, 2026-06-30

illo

Takeaway: Illo is now both a tool page and an imported executable skill at skills/engineering/illo/. The durable idea is repeatable visual identity: an agent skill with character packs, backends, references, quality checks, and scripts can make editorial images feel like they came from one brain instead of a pile of random generations. Source: X/@trevin, 2026-06-12; Source: illo skill docs, 2026-06-30

OrcaRouter

Takeaway: OrcaRouter is a programmable model-routing gateway, not an OpenRouter alias. Its Routing DSL turns model choice into versioned infrastructure: YAML + CEL rules over request shape, task classification, headers, tenant context, and agent session state, with fan-out, arbiters, shadow mode, canaries, and rollback. Keep the thought for LLM Council and local routing design. Source: X/@OrcaRouter, 2026-06-15; Source: OrcaRouter docs, 2026-06-30

Hermes Agent runtime cluster

Takeaway: Hermes Agent (Nous Research) is now promoted as a full runtime cluster, not just a /learn bookmark. The v0.17 Reach release makes Hermes a persistent personal-agent harness with messaging reach, background subagents, automation blueprints, skill hub upgrades, and memory improvements. /learn is the procedural-memory path: feed docs, directories, URLs, sessions, PDFs, or configs and let Hermes produce a standards-compatible SKILL.md. The memory provider and context engine plugin docs add the missing boundary: frequently reused knowledge should become memory/context infrastructure or wiki/capsule load rules, while repeatable procedures become skills. MoA is the model-routing/council surface inside Hermes, not a new skill. Source: X/@NousResearch, 2026-06-19; Source: X/@NousResearch, 2026-06-23; Source: X/@Teknium, 2026-06-23; Source: X/@Teknium, 2026-06-26; Source: X/@mr_r0b0t, 2026-06-26

The current version authority is GitHub, not the tweets: latest release is Hermes Agent v0.17.0 / tag v2026.6.19, published 2026-06-19; default branch main was checked at c9269fbfb6896793553b684eb6b5f3a79250fda7 on 2026-06-30 with no newer release tag found. The optional creative-ideation skill is a candidate, not an active local skill yet: it should be imported as a full folder with references, audited with skill-auditor, and tested against Kevin examples before promotion. Source: GitHub NousResearch/hermes-agent, 2026-06-30; Source: optional skill

Factory AutoWiki

Takeaway: Factory AutoWiki graduates from the coverage-floor table to a focused tool page and Agent-Docs Mesh comparison. The important idea is docs as a build artifact: run /wiki on demand, /install-wiki for push-based refresh, then expose generated repo docs through the Factory app, GitHub Wiki, Droid sessions, and optionally committed droid-wiki/ artifacts. The follow-up Open Source AutoWikis gallery proves the shape at corpus scale: 100 repos, 4.7k generated pages, 14 languages, 6.7M indexed stars, and generated page classes for architecture, systems, packages, contribution guides, lore, diagrams, video walkthroughs, and API maps. Use the pattern for generated architecture/module/flow docs, but keep Kevin-owned AGENTS.md, SKILL.md, fill slots, routes, and doctors as the operational source of truth. Source: X/@FactoryAI, 2026-06-17; Source: X/@FactoryAI, 2026-06-18; Source: Factory docs/Open Source AutoWikis, 2026-07-01

Re-sync 2026-06-29 (+new)

Voice agents on Vercel AI Gateway

Takeaway: Vercel AI SDK now has a voice-agent lane: realtime transport, speech generation, and transcription are surfaced through AI SDK 7 and AI Gateway. Use this when a TypeScript app needs voice IO with provider routing rather than a separate voice-only stack. Source: X/@vercel, 2026-06-29; Source: Vercel blog

Version check: npm ai latest is 7.0.11, @ai-sdk/react is 4.0.12, and @ai-sdk/workflow is 1.0.11 as of 2026-07-01. Check the exact subpackage before using a voice-agent primitive. Source: npm registry packument, 2026-07-01

Sakana Fugu Ultra on AI Gateway

Sakana Fugu Ultra is live on AI Gateway. Mythos-class intelligence in a single call, with a whole pool of models behind it. model: sakana/fugu-ultra - @vercel_dev | 2026-06-22 | 847 likes, 298 bookmarks

Takeaway: Promoted to Sakana Fugu Ultra. This is a hosted model-pool route, not an open checkpoint: Gateway exposes model: "sakana/fugu-ultra" while Sakana coordinates several frontier models behind the slug. Use it as AI Gateway orchestration evidence beside Vercel AI SDK, OpenRouter, and OrcaRouter, with no package semver beyond the checked Gateway model status. Source: X/@vercel_dev, 2026-06-22; Source: Vercel changelog/model page, checked 2026-07-04

Remote coding machines

Takeaway: VPS-hosted Claude Code/Codex is the raw form of Agent Machines: rent compute, set up dotfiles and packages, secure it with Tailscale/Cloudflare tunnels/firewall rules, and attach from any client. The wiki route remains managed persistent agent machines, but the bookmark validates the operator demand. Source: X/@levelsio, 2026-06-28; Source: X/@zack_overflow, 2026-06-28

Agent-company and KPI-agent workflows

Takeaway: Deep review promoted Greg's four diagrams into Agent Company OS as an operator checklist: humans keep strategy/taste/judgment, agents run bounded execution, the stack starts with clean data and structured knowledge, high-value opportunities are both repetitive and workflow-complex, and the best loops remove human maze steps while keeping policy checks and approval. Giga Scout adds the KPI-driven product version: set the metric, let agents test changes, learn from conversations, and improve the number. Source: X/@gregisenberg, 2026-06-27; Source: local artifact review, 2026-07-03; Source: X/@GigaAI, 2026-06-29

Agent eval and systems reading

Takeaway: Agentic loops are bounded by their evaluator. The Red Queen paper is now promoted into Red Queen Gödel Machine and the local implementation route is Agent Self-Improvement Eval Library: self-improvement needs receipt-based proof plus evaluator hardening, not a frozen judge. The Hitchhiker guide is now Hitchhiker's Guide to Agentic AI: a source map for model substrate, RL/reasoning, memory/RAG, harnesses, MCP/A2A, multi-agent systems, agentic UI, evaluation, and deployment. Source: X/@omarsar0, 2026-06-28; Source: X/@not_ellington, 2026-06-29

Skill-layout and research-loop fragments

Takeaway: Matt Pocock's .agents/skills symlink note reinforces the Cross-Agent Harness Portability runtime-index shape: keep one canonical skill corpus and project it into harness-specific folders. The proposed /research skill remains watch-only until the owner boundary is clear. Source: X/@mattpocockuk, 2026-06-28

Batch ledger

Takeaway: Full 23-record source ledger lives at X Bookmark Ingest 2026-06-29, including Vercel voice agents, remote coding machines, StyleX/design-system migration, Text-to-Lottie, Cloudflare app stack, Giga Scout, agent company diagrams, and lower-signal recruiting/dev records. Source: raw/x-bookmarks/enriched/, 2026-06-29

Re-sync 2026-06-28 (+new)

File-native agent directories

Takeaway: Omar Sar's "agent is a directory of files" bookmark reinforces Vercel Eve, Agent Legibility, Agent Harness, Runtime, Memory, and Evals, and the local agents/codex-self-improver/ contract: tools, skills, subagents, connections, schedules, evals, and package metadata should be inspectable on disk. This is the practical shape for Kevin's "brain as agent operating system" work. Source: X/@omarsar0, 2026-06-27

BINEVAL

Takeaway: BINEVAL gives the eval loop a concrete LLM-as-judge decomposition method: turn criteria into atomic yes/no questions, inspect verdicts, and aggregate scores. Use it when a holistic judge score is too opaque. Source: X/@omarsar0, 2026-06-27; Source: arXiv:2606.27226

OpenRouter open-weight agentic models

Takeaway: OpenRouter now treats open-weight models as serious agentic routing candidates: DeepSeek V4 Flash for cost/performance agentic coding, GLM-5.2 for open agentic coding portability, MiniMax M3 for multimodal long-context agents, and NVIDIA Nemotron 3 Ultra when U.S.-built fully-open deployment is part of the requirement. Claims stay eval-gated. Source: X/@OpenRouter, 2026-06-27; Source: OpenRouter Insights

Local coding agents

Takeaway: Local Coding Agents is now the dedicated route for open-weight/local agent stacks. The key decomposition is model -> local runtime -> coding harness; the adoption test is whether privacy, cost, modifiability, or learning is the blocker, with the same receipt fixture used against hosted Codex/Claude Code. Source: X/@rasbt, 2026-06-27; Source: Sebastian Raschka, 2026-06-27

Hermes Mixture of Agents 2.0

Takeaway: Hermes Agent (Nous Research) MoA is now a virtual model provider inside the runtime, preserving tools, transcript persistence, session context, interrupts, and normal Hermes surfaces. This is LLM Council as runtime plumbing. Source: X/@Teknium, 2026-06-26; Source: Hermes docs

Agent-browser durable sessions

Takeaway: Browser Testing Skills v0.31 adds durable worktree-scoped browser sessions, auto-restore, validation checks, transactional saves, and diagnostics. This sharpens the separate route for browser automation vs localhost verification vs search/discovery. Source: X/@ctatedev, 2026-06-27

Agentic marketing workflows

Takeaway: Promoted to Social Content OS, Frontend and Design Skills, Content Pipeline Workflow, and Loopy. The reviewed image is a no-slop loop, not just a prompt tip: context files for brand/product/ICP/offer, fresh monthly research from socials/blogs/YouTube//last30days, agent method proposal, two-model/two-pass review, taste/proof against experts and Kevin's own performing work, then ship and loop results back into research. Source: X/@EXM7777 and local image review, 2026-07-04

Batch ledger

Takeaway: Full same-day source ledger lives at X Bookmark Ingest 2026-06-28, including the initial 60-record pass plus the later 4-record delta: Next/Vercel repeats, Loopy, local coding agents, OpenRouter MCP, Rivet agentOS, ctx.ai, design-system signals, DeepSec, HyperFrames keyframing, agentic marketing workflow design, and awesome-evals thread fragments. Source: raw/x-bookmarks/enriched/, 2026-06-28

Re-sync 2026-06-26 (+new)

Next.js 16.3 agent routing shift

Takeaway: Next.js now encodes the official Next.js agent split: versioned local docs and managed AGENTS.md own framework knowledge; first-party skills own multi-step workflows such as next-dev-loop, next-cache-components-adoption, and next-cache-components-optimizer. This is a direct source for Kevin's own wiki/router distinction between reference pages and executable skills. Source: X/@nextjs, 2026-06-26; Source: Next.js blog, 2026-06-26

AI SDK Harness: Deep Agents and OpenCode

Takeaway: AI SDK HarnessAgent now supports five harness families: Claude Code, Codex, Deep Agents, OpenCode, and Pi. This strengthens the "harness portability" thesis: products can expose one AI SDK stream while choosing a specialized runtime behind it. Source: X/@vercel_dev, 2026-06-26; Source: Vercel changelog, 2026-06-25

Product design as an agent skill

Takeaway: Vercel's product-design system is a concrete pattern for Design Engineering and design-engineering-polish: one skill routes into canonical references, linters enforce deterministic rules, evals test whether guidance changes agent behavior, and a review loop turns accepted product decisions from Slack/Figma/GitHub into governed guidance. This is design taste as a repo-maintained operating system, not just a design blog post. Source: X/@vercel_dev, 2026-06-25; Source: Vercel blog, 2026-06-25

Open Tag

Takeaway: Open Tag is now promoted into Copilotkit, AG-UI Protocol, Agent GUIs, AI-Native Design Patterns, and Human-in-the-Loop Control. The product/API surface was still early access in the bookmark, so implementation claims stay version-checked; the durable pattern is cross-surface agent UI with generative UI, streaming replies, human approvals, and full thread context across Slack/Teams and future clients. Source: X/@ataiiam, 2026-06-26; Source: GitHub/npm, 2026-06-30

Evals source library

Takeaway: The BenchFlow awesome-evals collection is a source map for Eval-Driven Development and Agent Trajectory Evaluation: papers, blogs, talks, tools, and benchmarks for evaluating AI agents. The deep-reviewed screenshot thread now feeds Agent Self-Improvement Eval Library and agent-eval-library: choose graders by evaluated object, then route to deterministic assertions, trajectory/state checks, binary judges, benchmark-integrity checks, or evolving evaluators. Keep the thread collapsed into one evals source instead of ten separate bookmark notes. Source: X/@xdotli, 2026-06-24; Source: GitHub

Re-sync 2026-06-25 (+new)

Agent Company OS

Takeaway: Agent Company OS is the cleanest diagram in this batch: workspace brain + memory/skill system + runtime command plane + department leads + worker pool + proof loop. It updates Kevin's agent architecture pages because it explains how scoped workers stay useful without becoming blind, and how artifacts improve the memory/skill layer after each run. Source: X/@DerekNee, 2026-06-25

Resolvers: routing table for intelligence

Takeaway: Promoted through The Resolver Pattern and Knowledge Brain System. Garry Tan's article is the external write-up for Kevin's resolver architecture: skills are capabilities, but resolvers make them reachable, fileable, and auditable. Source review recovered the article through X metadata and a readable Glean mirror because anonymous X Article fetch returned only a login shell. Source: X/@garrytan and Glean mirror, 2026-07-04

Latent Briefing

Takeaway: Promoted to Latent Briefing and Agent Memory System Architecture. Ramp Labs' X Article announces representation-level context sharing for multi-agent systems: compact task-relevant KV-cache state from the orchestrator's reasoning instead of sending the whole text trace to every worker. This is an inference/memory architecture signal, not a package-backed dev tool. Source: X/@RampLabs and technical source review, 2026-07-04

Matt Pocock writing skills

Takeaway: Matt Pocock's Skills (Skills for Real Engineers) now contributes a local long-form writing pipeline: Writing and Content Skills for exploration, Writing and Content Skills for grounded beat-by-beat drafting, and Writing and Content Skills for direct paragraph/block shaping. This complements social-draft instead of replacing it: article first, platform copy later. Source: X/@mattpocockuk, 2026-06-25; Source: GitHub in-progress skills

BuilderIO /visual-plan

Takeaway: BuilderIO Skills is a candidate source for visual planning and recap artifacts. The /visual-plan demo matters because user-flow review should be a storyboard or interactive plan when a chat paragraph would hide too much product shape. Keep reference-first until the hosted/local Agent-Native Plan dependency is audited. Source: X/@Steve8708, 2026-06-25; Source: BuilderIO/skills

Hermes /learn explainer

Takeaway: Hermes Agent (Nous Research) /learn is the self-improvement loop in product form: feed source material, sessions, docs, URLs, PDFs, directories, or configs, then crystallize a tested reusable skill before the painful session has to repeat. This bookmark is an explanatory reinforcement of the existing Hermes /learn page, not a new tool. Source: X/@akshay_pachaar, 2026-06-25

AI SDK 7 control surfaces

Takeaway: Vercel AI SDK is now a wider agent-app surface: reasoning control, tool approvals, tool/runtime context, file and skill uploads, MCP Apps, durable workflows, terminal UI, sandbox support, harness integrations, telemetry, lifecycle events, and migration support. The local ai-sdk skill now routes by primitive instead of assuming one loop pattern: WorkflowAgent for durable workflows, AI SDK HarnessAgent for mature harness embedding, and lower-level Agent / streamText / generateText for app-owned loops. Source: X/@aisdk, 2026-06-25; Source: AI SDK v7 WorkflowAgent docs

AI engineer stack: harness > prompt

Takeaway: The @divaagurlxw checklist reinforces Harness Engineering: AI engineers should learn harness engineering, context engineering, caching tradeoffs, KV-cache behavior, prefill vs decode latency, tool protocol costs, evals, and observability rather than stopping at prompt phrasing. Deep review promoted the linked distributed-systems mapping into Agent Resilience Patterns: circuit breakers become kill switches, bulkheads become blast-radius limits, sagas become compensation chains, and traces become model/tool/workspace action traces. Source: X/@divaagurlxw, 2026-06-24

Claude token grey-market signal

Takeaway: The HN-linked bookmark about discounted Claude-token resale is not a "buy this" signal; it is a Agent Unit Economics and Agent Security Model signal. The reviewed HN screenshot made the risk concrete: below-list access can be tied to account pooling, suspicious payments, and possible output/reasoning-trace resale. If agent tokens have a grey market, products need provider provenance, billing anomaly checks, credential hygiene, and spend controls instead of treating cheap tokens as a harmless optimization. Source: X/@GregKamradt, 2026-06-25; Source: Hacker News thread, 2026-06-25

Origami outbound agents

Takeaway: Origami is a reference for agentic GTM surfaces: type a domain, infer an ideal-customer profile, and run customer-discovery/outreach agents. Keep it reference-only until deliverability, consent, rate limiting, and provenance are verified. Source: X/@origamichat, 2026-06-25; Source: origami.chat

Matt Pocock /loop-me

Takeaway: Frontend and Design Skills is installed locally as the interview skill for finding recurring day-to-day work Kevin should delegate to agents. It is the no-one-off rule with a microphone: identify loops, write specs, then promote to skills or automations. Deep review refreshed mattpocock/skills to main 0877403d1e867fd9d574117e9b34ade404f36d2a and kept the Kevin-owned local adaptation rather than importing a duplicate upstream router. Source: X/@mattpocockuk, 2026-06-24; Source: GitHub SKILL.md

Matt Pocock no-op skill pruning

Takeaway: Browser Testing Skills now treats "tighten/prune this skill" as a first-class branch. The rule is contextual deletion testing: if removing a line does not change the agent's behavior for that skill, delete it or rewrite it as a checkable branch, artifact, command, threshold, or done condition. Source: X/@mattpocockuk, 2026-06-24; Source: X/@joenandez, 2026-06-25

Agent memory is a data system

Takeaway: Are We Ready For An Agent-Native Memory System? and MemoryData reframe agent memory as a data-management system with module-level evaluation: representation/storage, extraction, retrieval/routing, and maintenance. This strengthens Kevin's Agent Memory Patterns and wiki-as-memory architecture. Source: X/@dair_ai, 2026-06-24; Source: arXiv 2606.24775

Hermes creative ideation skill

Takeaway: Hermes Agent (Nous Research) added an optional creative-ideation skill that analyzes a prompt and routes through 22 creative methodologies. This is a useful pattern for Kevin's local skills too: encode creative methods as a selectable procedure, then test whether method routing beats generic brainstorming. Source: X/@NousResearch, 2026-06-24; Source: GitHub skill

OpenRouter Image API

Takeaway: OpenRouter Image API makes image generation more agent-addressable: /api/v1/images/models exposes typed capabilities across providers, a normalized request shape, streaming support for compatible models, and exact USD cost reporting. Deep review refreshed the source snapshot: 38 image models across nine provider namespaces on 2026-06-30, with endpoint-level pricing, passthrough, and capability descriptors documented for agent-side validation. Source: X/@OpenRouter, 2026-06-24; Source: OpenRouter docs/API, 2026-06-30

Re-sync 2026-06-24 (+today-stamped)

Hermes /learn — source corpus → tested skill (NousResearch / Teknium)

Takeaway: Hermes Agent (Nous Research) now has /learn: feed it directories, sessions, docs, manuals, PDFs, configs, URLs, or other source material and it distills a reusable skill. This is a direct accelerator for Kevin's no-one-off workflow, but still needs the same acceptance bar: tested SKILL.md, not a vibes-only prompt. Source: X/@NousResearch, 2026-06-24; Source: X/@Teknium, 2026-06-24

Hermes config defaults

Takeaway: Hermes Agent (Nous Research) now records the config-default/token-budget lesson from @IBuzovskyi's artifact. Official docs make agent.reasoning_effort the clearest verified knob: unset means medium; valid values are none, minimal, low, medium, high, and xhigh; higher values trade latency/tokens for deeper reasoning on hard tasks. The reviewed video proves Hermes setup/config UI exists, but not all eight claimed config values, so exact settings stay docs-backed. Source: X/@IBuzovskyi and local video review, 2026-07-04; Source: Hermes configuration docs

Agent command surfaces — Linzumi + GUIs for agents

Takeaway: Linzumi and Agent GUIs point to the same product thesis: once agents are numerous, terminals and isolated IDE chats stop being enough. Teams need shared command rooms, decision inboxes, resumable state, review artifacts, and unblock loops. The 2026-07-03 deep review preserved the local video evidence: GitHub/request cards, task status dots, and an approval card that lets the human keep the current plan or redirect the agent. Source: X/@ycombinator and local video review, 2026-07-03; Source: X/@Tocelot, 2026-06-24

Durable agent sessions

Takeaway: OpenComputer Durable Agent Sessions is another data point for Persistent Sandbox Thesis: agent work needs durable execution state, not just chat history. Source: X/@utpalnadiger, 2026-06-24

Batch ledger + artifacts

Takeaway: Full 40-record artifact ledger lives at X Bookmark Ingest - 2026-06-24, including original tweet embeds, relevant replies, resolved links, and local media.

Re-sync 2026-06-21 (+new)

Open Knowledge Format (OKF) — Google Cloud (GoogleCloudTech)

Introducing the Open Knowledge Format (OKF), an open specification that formalizes the LLM-wiki pattern into a portable, interoperable format. - @GoogleCloudTech | 2026-06-16 | 5,298 bookmarks

Takeaway: Open Knowledge Format (OKF) — vendor-neutral spec (v0.1, 2026-06-12) that standardizes the exact Karpathy LLM-wiki pattern this wiki runs on (markdown + frontmatter, one concept/file, cross-links as graph, reserved index.md/log.md, only type required). Deep image review confirms the concrete file shape: YAML frontmatter, Markdown schema table, and link-based joins. kevin-wiki is already an OKF bundle in all but field-naming. Source: X/@GoogleCloudTech, 2026-06-16

Codex Record & Replay — demo→skill (OpenAIDevs)

Show Codex a workflow once. Reuse it as a skill. Record & Replay lets you show Codex a recurring task … Codex turns that demo into an inspectable, editable skill. - @OpenAIDevs | 2026-06-18 | 9,595 bookmarks

Takeaway: Codex Record & Replay — the No One-Off Work principle as a product feature (macOS, build 26.616, needs Computer Use; excludes EEA/UK/CH). Dedicated page created. Source: X/@OpenAIDevs, 2026-06-18

Cursor Origin + SpaceX acquisition (cursor_ai / SpaceX)

We're launching code storage and git hosting. Origin gives teams and agents a place to host, review, and collaborate on code. - @cursor_ai | 2026-06-16

Takeaway: Cursor Origin — git hosting built for agents as committers. Source-reviewed 2026-07-04: the live waitlist page frames Origin as a git forge for agent-era code velocity, but still exposes only a hosted waitlist rather than a public package/API version. Launched the same day @SpaceX announced an all-stock acquisition of Cursor + a jointly-trained model for Cursor/Grok Build. Dedicated page refreshed; Cursor 2.0 and Composer already tracks the joint-model context. Source: X/@cursor_ai, 2026-06-16; Source: Cursor Origin waitlist, 2026-07-04; Source: X/@SpaceX, 2026-06-16

Loop Library — curated agent loops (MatthewBerman)

Just launched Loop Library - a curated list of agent loops you can use right now. - @MatthewBerman | 2026-06-18 | 7,267 bookmarks

Takeaway: Loop Library now has 70 published loops in the live catalog, updated 2026-06-26, plus agent-readable catalog.json, catalog.md, catalog.txt, llms.txt, and /agents/ surfaces. The repo authority is Loopy (Forward-Future/loopy), which owns finding, adapting, auditing, running, debriefing, and publishing bounded loops. Maps to Borrowed Intelligence (Drain It into Plans) / The Eval Loop (Slop Is an Output Problem), but remains eval-gated before adoption. Source: X/@MatthewBerman, 2026-06-18; Source: Loop Library catalog and GitHub, 2026-07-03

Unreal Engine 5.8 — experimental MCP server (UnrealEngine)

Unreal Engine 5.8 ships today with experimental MCP server support … configure the MCP plugin and connect to any agent. - @UnrealEngine | 2026-06-17 | 4,765 bookmarks

Takeaway: MCP reaches the creative-tooling tier — evidence for MCP as the Integration Standard as the default agent-integration surface beyond dev tools. Deep review added Epic's experimental/local/no-auth/game-thread caveats and preserved the PCG editing artifact. Source: X/@UnrealEngine, 2026-06-17; Source: Epic Unreal MCP docs, 2026-07-02

Photon Spectrum — iMessage App API for agents (daniel_tian1)

We're launching the world's 1st iMessage App API for agents. your agents can now send Apps inside iMessage — an Uber Eats menu, a mini game, or any agentic UI. - @daniel_tian1 | 2026-06-17 | 2,795 bookmarks

Takeaway: Photon Spectrum — TypeScript framework (spectrum-ts@8.1.1) for bringing agents to iMessage, WhatsApp, Telegram, Slack, terminal, and custom messaging providers. Local media shows a rich iMessage app surface, not just text chat. Promoted from hub-only because the repo/package is public and versioned. Source: X/@daniel_tian1, 2026-06-17; Source: npm registry packument, 2026-07-01

Palmier — video editor built for AI (Marcos12345rico)

today we're launching @Palmier_io, a video editor Claude can edit. use AI to edit, organize, and generate footage directly in the timeline. open-source. mac native. - @Marcos12345rico | 2026-06-17 | 14,295 bookmarks

Takeaway: Palmier — Mac-native AI video editor with an HTTP MCP server (http://127.0.0.1:19789/mcp) for Claude/Codex/Cursor timeline editing. Promoted from hub-only after repo/version review (palmier-pro@v0.4.5, GPL-3.0, 9.6K stars). Source: X/@Marcos12345rico, 2026-06-17

PixelRAG — visual retrieval without HTML parsing (akshay_pachaar)

Web scraping will never be the same. (100% open-source visual search at scale) PixelRAG … skips HTML parsing completely. - @akshay_pachaar | 2026-06-20

Takeaway: PixelRAG — retrieves over rendered page pixels instead of scraped-text embeddings; a different point on the Retrieval Quality spectrum and relevant to agent web-reading/scraping. Deep review promoted it from hub-only after primary-source/version checks: Apache-2.0, pixelrag==0.3.0, release v0.3.0, hosted API over 8.28M Wikipedia pages, pixelshot CLI, and a Claude Code pixelbrowse plugin path. The boundary is also now explicit: auth, collapsed content, paywalls, captchas, and render timing still need browser automation discipline. Source: X/@akshay_pachaar, 2026-06-20; Source: StarTrail-org/PixelRAG, 2026-06-30

Re-sync 2026-06-17 (+new)

eve — Vercel's agent framework (vercel)

Introducing eve, an agent framework. agent/ agent.ts instructions.md tools/ skills/ sandbox/ schedules/ … Like Next.js, for agents. - @vercel | 2026-06-17

Takeaway: Vercel Eve — filesystem-first; durable execution + sandbox + HITL approvals + evals built in. "An agent is a directory of files" mirrors Kevin's Agent-Docs Mesh; overlaps Dedalus Labs agent-compute. Dedicated page created. Source: X/@vercel, 2026-06-17

GLM-5.2 — open-weights long-horizon coding model (Zai_org)

Introducing GLM-5.2: Frontier Intelligence, Open Weights — significant coding/agentic gains, 1M context, two reasoning-effort levels (max / high). - @Zai_org | 2026-06-17

Takeaway: GLM-5.2 — strongest open-source coding model at launch (~753B MoE, ~40B active). Deep review added the launch-image caveat: Z.ai's chart shows GLM-5.2 score rising with effort level, but max effort costs far more output tokens, so route it as an effort/cost tradeoff rather than a free win. Theo's later reviewed DeepSWE cost plot sharpened that caveat: glm-5.2 [max] is not automatically cheaper once output-token volume and wall-clock latency are counted, even when raw tokens are cheaper. On Vercel AI Gateway as zai/glm-5.2, the official changelog confirms 1M context, AI SDK usage, retries/failover, ZDR, key budgets, BYOK, and no platform markup; the live Gateway model page now exposes per-provider routing data. The @vercel_dev self-thread adds Next.js eval evidence: GLM 5.2 via OpenCode scored 88% base and 96% with AGENTS.md on the 2026-06-09 run. Ollama cloud, Design Arena, FrontierSWE, Theo's DeepSWE cost plot, and BridgeBench remain the other dated benchmark anchors. ZCode is z.ai's GLM-5.2 coding agent. Source: X/@Zai_org, 2026-06-17; Source: X/@vercel_dev, 2026-06-16; Source: Vercel changelog, checked 2026-07-04; Source: Next.js evals, checked 2026-07-04; Source: X/@ProximalHQ, 2026-06-16; Source: X/@Designarena, 2026-06-16; Source: X/@theo, 2026-06-21; Source: X/@bridgemindai, 2026-06-14

OpenRouter Fusion — compound model route (OpenRouter)

Introducing the Fusion API, the smartest compound model in the market. Fusion achieves Fable-level intelligence at half the price. - @OpenRouter | 2026-06-13 | 14,782 likes, 13,438 bookmarks

Takeaway: OpenRouter Fusion — panel + judge synthesis exposed as a model alias, server tool, and plugin. Use for deep research or architecture critique, not routine coding. Deep review on 2026-07-02 checked the live docs/model page and local chart artifact: Fusion is still a multi-model deliberation pipeline with Quality/Budget presets and overrideable panel/judge fields, while the chart remains a dated benchmark snapshot. Source: X/@OpenRouter thread, 2026-06-13; Source: OpenRouter docs/model page, 2026-07-02

Re-sync 2026-06-13 (+9 new)

AI SDK HarnessAgent (vercel_dev)

AI SDK now supports agent harnesses like Claude Code, Codex, and Pi with sandboxed sessions and AI SDK-compatible streams: const agent = new HarnessAgent({ harness: claudeCode, sandbox: … }) - @vercel_dev | 2026-06-12 | 635 likes, 546 bookmarks

Takeaway: AI SDK HarnessAgent — swap the harness like you swap a model; official @vercel_dev post for the Harness Engineering launch. Deep review on 2026-07-04 refreshed the initial adapter snapshot: Claude Code 1.0.18, Codex 1.0.19, Pi 1.0.18, and Vercel Sandbox 1.0.18. Source: X/@vercel_dev, 2026-06-12; Source: npm registry, 2026-07-04

Hermes Automation Blueprints (NousResearch)

Hermes Agent now has Automation Blueprints, turning cron jobs into clickable, fillable, conversational workflows. - @NousResearch | 2026-06-11 | 2,733 likes, 1,472 bookmarks

Takeaway: Hermes Agent (Nous Research) now treats repeatable agent work as a template surface, not just raw cron. The video artifact shows "a blueprint for every routine" and "one function for every path": fillable forms sit in front of schedule, webhook/API, and GitHub-event triggers so recurring workflows become operator-friendly products. The docs-only follow-up (2065200919038292326) confirms the current Automation Blueprints guide as source authority for Schedule, GitHub Event, and API Call triggers, plus delivery to Telegram, Discord, Slack, SMS, email, GitHub comments, or local files. This maps directly to Kevin's automations/ model: PR review, backlog triage, docs drift, bookmark ingest, and source freshness should expose form-like launch controls while keeping the underlying schedule/versioning in files. Source: Hermes docs, checked 2026-07-04; Source: X/@NousResearch, 2026-06-11; Source: X/@NousResearch docs follow-up, 2026-06-11

Claude Fable 5 + HyperFrames autonomous video (nateherk)

Claude Fable 5 made this entire video by itself. I gave it a /goal prompt, went to the gym, and came back to this. Even the sound effects. Shoutout @HyperFrames_ - @nateherk | 2026-06-12 | 3,658 likes, 4,091 bookmarks

Takeaway: HyperFrames + Claude Fable 5 — /goal -> finished video unattended; autonomous agent-video pipeline proof. The reviewed frames show a synthetic/cloned talking-head sequence and a final prompt surface, so this is capability evidence for an unattended media loop, not a new default production dependency. Source: X/@nateherk, 2026-06-12

Batch 26-150 (2026-06-13)

Karpathy wiki DIY stack (NickSpisak_)

Made an updated version this weekend... Grab @karpathy's latest gist... qmd... Setup a CLI called "brain" - @NickSpisak_ | 2026-04-06 | 3,574 likes, 9,339 bookmarks

Takeaway: DIY Karpathy LLM Wiki Setup Guide pipeline (gist + summarize + yt-dlp + Obsidian + QMD - Local Wiki Search Engine) wrapped in a local brain CLI. Deep artifact review confirmed the visual target is an inspectable graph view with filters and concept clusters, so the reusable idea is "make the memory navigable," not a new dependency. Source: X/@NickSpisak_, 2026-04-06

Anthropic employee wedding site via Claude Code

I got married... had Claude Code analyze 12 years of iMessages with my wife, then Claude Design whipped up a website... - @helloitsaustin | 2026-05-05 | 18,941 likes, 8,167 bookmarks

Takeaway: Promoted to AI-Native Design Patterns as a private-archive-to-data-story pattern. The reviewed images show aggregate iMessage history turned into a wedding-site narrative with counts, yearly rhythm, hourly heatmap, love-you count, and emoji vocabulary; the durable lesson is consented, aggregate personal data storytelling, not a new tool install route. Source: X/@helloitsaustin, 2026-05-05; Source: local image artifact review, 2026-07-02

Ernesto vibecoding startup prompt

Vibecoding a startup without sending this prompt first, is a waste of time. - @ErnestoSOFTWARE | 2026-03-17 | 4,617 bookmarks

Takeaway: Promoted to Vibecoding Critique as the structural-audit preflight variant: dead code, folder structure, hardcoded values, naming, scalability risks, worst-file rewrite, and README documentation before a vibe-coded prototype is treated as a serious project. Source: X/@ErnestoSOFTWARE and local artifact review, 2026-07-03

Colgate LLM synthetic panels

A toothpaste company has quietly killed the entire market research industry... Colgate published a paper showing you can predict real purchase intent... - @HowToAI_ | 2026-06-11 | 7,304 likes, 11,375 bookmarks

Takeaway: LLM roleplay panels hit ~90% of human test-retest reliability on purchase-intent surveys (not 90% absolute accuracy). Deep review verified arXiv v3 and the paper screenshot, then promoted the stronger evaluator pattern: elicit qualitative text first, score second with SSR, and treat synthetic panels as screening rather than proof of market demand. See LLM Synthetic Panels (Purchase Intent). Source: X/@HowToAI_ thread, 2026-06-11; Source: arXiv 2510.08338, 2025-10-27

Boris Cherny — write loops not prompts

Claude Code creator: "I don't prompt Claude anymore. I write loops - and the loops do the work." - @0xCodez | 2026-06-04 | 3,982 likes, 10,276 bookmarks

Takeaway: Deep review promoted this into Agent Looping: the durable unit is the loop that reads context, runs tools, verifies, and repeats under constraints. This is not a new skill because loopy, agent-iteration-loop, and agent-eval-library already own Kevin's executable loop routes. Source: X/@0xCodez, 2026-06-04; Source: WorkOS YouTube

Claude Code Workflows (trq212)

Workflows are the biggest upgrade to Claude Code's capabilities since skills and subagents. - @trq212 | 2026-06-02 | 4,690 likes, 8,217 bookmarks

Takeaway: Source-reviewed against the current Claude blog and docs, then folded into Claude Code Dynamic Workflows. Dynamic workflows are generated JavaScript orchestration scripts for high-value tasks that need fan-out, saved rerunnable coordination, or adversarial review; /workflows exposes run/cost state, and Claude Code v2.1.154+ is the documented floor. No local media was expected; the blocked X photo reply is not source-critical. Source: X/@trq212, 2026-06-02; Source: Claude Code docs, reviewed 2026-07-01

Fable 5 — audit existing projects first

Got your hands on Claude Fable 5? The first thing you should do is to upgrade your main projects with it... - @meta_alchemist | 2026-06-09 | 3,703 likes, 10,417 bookmarks

Takeaway: Resolved artifact points to Andrej Karpathy's LLM Coding Skills, not a generic Fable upgrade recipe. Keep as source evidence for assumptions, simplicity, surgical diffs, and verifiable success criteria; do not import a duplicate local rule file. Source: X/@meta_alchemist, 2026-06-09

AutoResearch for marketing

Karpathy's AutoResearch is changing how campaigns get optimized... Ole Lehmann tested landing page copy, 56% → 92% pass rate overnight. - @shannholmberg | 2026-03-24 | 3,792 likes, 10,402 bookmarks

Takeaway: AutoResearch (Karpathy) - Applied to Marketing is the durable route. Deep review adds the actual visual pattern: write a 3-6 item checklist, run the skill once, let the loop mutate one prompt rule, score it, keep if better, revert if worse, and save a changelog. Current source authority is untagged karpathy/autoresearch master at 228791f, 89,666 stars, no public release tags. Source: X/@shannholmberg, 2026-03-24; Source: GitHub karpathy/autoresearch, 2026-07-03

Matt Van Horn — Printing Press stack thread

(X article on daily agent stack) - @mvanhorn | 2026-06-02 | 3,243 likes, 9,823 bookmarks

Takeaway: Expanded beyond Printing Press - Agent-Native CLIs from a Single Prompt into AgentMail as a concrete email-to-Claude-Code gateway pattern. The X article body remains login-wall blocked, but the self-thread preserves the operator stance: Claude as the brain, Codex as the muscle, terminal sessions preferred, mobile Claude Code used daily, and high monthly overage accepted for throughput. The linked source repo shows the repeatable implementation: AgentMail or Primitive inbox -> allowlist and DKIM/SPF checks -> prompt file with attachment paths -> cmux/Ghostty starts a fresh Claude Code session. Source: X/@mvanhorn thread, 2026-06-02; Source: GitHub mvanhorn/agentmail-to-claude-code, 2026-07-02

Harness engineering checklist

As an AI Engineer. Please learn >Harness engineering, not just prompt engineering... - @divaagurlxw | 2026-06-04 | 4,277 likes, 7,184 bookmarks

Takeaway: Promoted into Harness Engineering as the 2026 AI-engineer skill map: harness engineering over prompt engineering, context engineering over long prompts, prompt caching versus semantic caching, KV-cache behavior, latency, tool overhead, evals, and observability. The self-thread adds only a notification request, so no separate artifact page was needed. Source: X/@divaagurlxw, 2026-06-04; Source: raw/x-bookmarks/enriched/2062419864908951606.json, reviewed 2026-07-04

32 Claude shortcuts (RoundtableSpace)

32 CLAUDE SHORTCUTS THAT MAKE YOUR PROMPTS 10X BETTER... - @RoundtableSpace | 2026-04-06 | 6,061 bookmarks

Takeaway: Promoted into Claude Prompting Principles as prompt vocabulary, not real built-in Claude shortcuts. The reviewed Google Sheets card is useful for naming common moves (summarize, checklist, role/task/format, schema, compare, bias check, self-eval, parallel lenses), but Kevin's harness should encode these as structure and skills rather than copy-pasted slash incantations. Source: X/@RoundtableSpace, 2026-04-06; Source: local image review, 2026-07-03

Open Higgsfield AI (self-hosted)

I was paying for Higgsfield AI. Then I found this repo and cancelled my subscription... - @ihtesham2005 | 2026-04-07 | 2,691 likes, 5,920 bookmarks

Takeaway: Deep review keeps Open Higgsfield AI as a skeptical self-hosted creative-UI reference: the viral 200+ model claim is not verified in the README, the repo description only claims 20+ models, GitHub API did not detect a license, and generation remains paused until provider credits are added. Source: X/@ihtesham2005 and GitHub review, 2026-07-04

How LLMs work (plain language)

LLMs explained without all that yucky math stuff - @0xkato | 2026-06-04 | 5,007 bookmarks

Takeaway: Educational article linked in tweet; onboarding reference. Source: X/@0xkato, 2026-06-04

Deep review: Promoted to LLM Transformer Machinery. The article's durable value is the ordered stack: tokenization, embeddings, positional encoding, attention, multi-head attention, FFNs/MoE, residual streams, normalization, and next-token decoding. Use it to ground model-behavior explanations in a concrete transformer surface instead of vague "the model knows" language. Source: 0xkato article, 2026-07-03

Model & Platform Announcements

Claude Managed Agents launch

Introducing Claude Managed Agents: everything you need to build and deploy agents at scale. It pairs an agent harness tuned for performance with production infrastructure, so you can go from prototype to launch in days. - @claudeai | 2026-04-08 | 56,902 likes

Takeaway: Claude Managed Agents is now a hosted agent control plane, not just a launch claim: the current docs model the surface as agent, environment, session, and events, with built-in bash/file/web/MCP tools, SSE streaming, stateful sessions, a required managed-agents-2026-04-01 beta header, and token pricing plus $0.08 per running session-hour. Deep video review showed the real product shape: managed workspaces, prompt templates, timeline bars, transcript/debug tabs, role tabs, and subagent creation. Source: X/@claudeai, 2026-04-08; Source: Anthropic Managed Agents docs, 2026-07-01

Claude Fable 5 and Mythos 5

Introducing Claude Fable 5: a Mythos-class model that we've made safe for general use. Its capabilities exceed those of any model we've ever made generally available. - @claudeai | 2026-06-09 | 104,049 likes | 21,895 bookmarks

Takeaway: Fable 5 is GA with topic fallbacks to Opus 4.8 on sensitive domains (~5% of sessions). Mythos 5 (fewer safeguards) restricted to Project Glasswing partners until trusted-access expands. Demo video in thread. Source: X/@claudeai thread, 2026-06-09

Project Glasswing - Anthropic security vulnerability scanning

Introducing Project Glasswing: an urgent initiative to help secure the world's most critical software. Powered by Claude Mythos Preview, which can find software vulnerabilities better than all but the most skilled humans. - @AnthropicAI | 2026-04-07 | 42,231 likes

Takeaway: Mythos Preview found zero-days in major OSes and browsers; $100M defensive credits. Deep source review keeps this in Project Glasswing as restricted defensive-deployment evidence, not a generally routable model. The system card and source thread are preserved there. Source: X/@AnthropicAI thread, 2026-04-07; Source: Anthropic Glasswing, 2026-04-07

Claude-Mem - Hook-based session memory

Observer hooks capture tool usage and replay semantic summaries across Claude Code sessions. - @tom_doerr | 2026-04-10

Takeaway: RAD-style working memory for agents. See Claude-Mem and Claude Code Harness. Source: compiled from enrichment, 2026-06-01

Glass - Meeting desktop agent

Invisible desktop assistant: proactive meeting notes and screen-aware answers. - @tom_doerr | 2026-04-15 | 415 bookmarks

Takeaway: Glass (Pickle) overlaps the meeting-agent category with cloud factories, but its useful angle is local-first desktop context: Electron/Next.js/Tailwind, screen/audio-aware answers, real-time notes, Ollama/Whisper path, and invisibility to recordings. Current review refreshed GPL-3.0, release v0.2.4, 7,547 stars, and CheatingDaddy fork provenance. Treat it as desktop-agent prior art unless security/license review passes. Source: X/@tom_doerr, 2026-04-15; Source: GitHub pickle-com/glass, 2026-06-30

Google Stitch AI-native design agent

Meet the new Stitch: AI-Native Canvas, Smarter Design Agent, Voice, Instant Prototypes, Design Systems and DESIGN.md. - @stitchbygoogle | 2026-03-18 | 40,823 likes

Takeaway: Google enters agent-native design with DESIGN.md, but the deep review makes this broader than a design-file export. Stitch is now an AI-native canvas with a whole-project design agent, voice critique/updates, instant prototypes, DESIGN.md import/export, MCP, SDK, and Stitch skills. The canonical URL is stitch.withgoogle.com; the old Labs path 404s. See Google Stitch. Source: X/@stitchbygoogle, 2026-03-18

ChatGPT Images 2.0

A state-of-the-art image model that can take on complex visual tasks and produce precise, immediately usable visuals, with sharper editing, richer layouts, and thinking-level intelligence. - @OpenAI | 2026-04-21 | 25,198 likes

Takeaway: OpenAI upgrades image generation with thinking-level intelligence and more reliable text/object composition. Deep media review found the launch video demonstrating a chameleon prompt across title/card/newspaper/scene variants, so route it to design/media-generation workflows that need legible text, controlled aspect ratios, and self-checked visuals. See ChatGPT Images 2.0 (gpt-image-2). Source: X/@OpenAI, 2026-04-21

Codex subagents

Subagents are now available in Codex. Spin up specialized agents to keep your main context window clean, tackle different parts of a task in parallel, steer individual agents as work unfolds. - @OpenAIDevs | 2026-03-16 | 8,001 likes

Takeaway: OpenAI Codex gets parallel subagent spawning - mirrors Cursor Orchestrate Skill pattern. Source: X/@OpenAIDevs, 2026-03-16

Cursor SDK announcement

We're introducing the Cursor SDK so you can build agents with the same runtime, harness, and models that power Cursor. Run agents from CI/CD pipelines, create automations for end-to-end workflows. - @cursor_ai | 2026-04-29 | 8,750 likes

Takeaway: Cursor exposes its agent runtime as an SDK - enables programmatic agent orchestration outside the IDE. Source: X/@cursor_ai, 2026-04-29

Cursor Design Mode

Use Design Mode in Cursor 3 to annotate and target UI elements in the browser. - @cursor_ai | 2026-04-07 | 1,673 likes

Takeaway: Cursor Design Mode gives the agent element identity plus screenshot/annotation context from the live browser. Deep video review confirms the loop: click target, element-context popover, selected target inserted into chat, text selection, freehand annotation, and region markup before edits. Source: X/@cursor_ai, 2026-04-07

GPT Realtime 2.0 startup ideas

GPT Realtime 2.0 is pretty incredible. 17 startup ideas that ONLY work because of what this model makes possible: 1. Real-time contract negotiation agent. Sits on a call between two parties, checks pricing tools and compliance databases in parallel, and suggests terms. - @gregisenberg | 2026-05-31 | 1,067 likes, 1,765 bookmarks

Takeaway: Promoted through Greg Isenberg. OpenAI's official GPT-Realtime-2 launch makes the capability concrete: live voice interactions with reasoning, interruption recovery, parallel tool calls, 128K context, and adjustable reasoning effort. The product pattern is Services-as-Software over voice: agents can do work during the call instead of merely summarizing it afterward. Source: X/@gregisenberg, 2026-05-31; Source: OpenAI, 2026-05-07

Agent Knowledge & Memory

Karpathy LLM Knowledge Bases

Something I'm finding very useful recently: using LLMs to build personal knowledge bases for several topics of research interest. A large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge. - @karpathy | 2026-04-02 | 54,109 likes

Takeaway: Karpathy validates the LLM Wiki pattern — this wiki's foundational architecture. See Karpathy LLM Wiki Setup Guide. 98,708 bookmarks (#1 in corpus). Source: X/@karpathy, 2026-04-02

Claude Code memory architecture (reverse-engineered)

Based on everything explored in the source code, here's the full technical recipe behind Claude Code's memory architecture. It isn't "store everything" but constrained, structured. - @himanshustwts | 2026-03-31 | 6,280 likes

Takeaway: Promoted through Claude Code Memory Architecture (Reverse-Engineered) and Claude Code Harness. Deep visual review confirms the three-temperature memory model: always-injected MEMORY.md, on-demand topic files, and grep-only JSONL transcripts, with manual writes, per-turn extraction, background autoDream, and consolidationLock.ts around background consolidation. Source: X/@himanshustwts, 2026-03-31; Source: local image artifact, 2026-07-03

Hermes Agent + LLM-Wiki

Hermes Agent now comes packaged with Karpathy's LLM-Wiki for creating knowledgebases and research vaults with Obsidian! - @Teknium | 2026-04-07 | 3,505 likes

Takeaway: Open-source agent ships with the LLM-Wiki pattern as a bundled skill - the pattern is becoming standard agent infrastructure. Deep review updated Hermes Agent (Nous Research) with the llm-wiki skill docs, graph artifact, and current v0.18.0 release snapshot. Source: X/@Teknium, 2026-04-07; Source: Hermes LLM-Wiki docs, 2026-07-02

"Infinite memory" context tool (lean-ctx signal)

YOU CAN NOW GIVE YOUR CLAUDE CODE INFINITE MEMORY FOR FREE - Up to 95% fewer tokens per session. 20× more tool calls before context limits. 100% open-source. - @RoundtableSpace | 2026-03-15 | 1,407 likes

Takeaway: Community validation of Lean Ctx thesis - massive token savings with context management. Deep review updated Claude-Mem, and the 2026-07-01 registry refresh moved the current package to claude-mem@13.9.2 with default HEAD cbdce2d676a9646b73dc236eebb53e4019870dcf; keep it reference-only until audited. Source: X/@RoundtableSpace, 2026-03-15; Source: npm registry packument, 2026-07-01

GBrain + Wiki retrieval analysis

LLM-Wiki hits a wall when you use it as a second brain for your agents: file-based retrieval breaks down fast. 10,000 lines in and grep becomes a liability. @garrytan just shipped GBrain and solved this elegantly. - @JoshuaIPark | 2026-04-10 | 257 likes

Takeaway: Promoted into Knowledge Brain System and Agent Search and Retrieval. The reviewed self-thread turns this from a vague "grep bad" take into a retrieval boundary: keep markdown/git as the canonical editable memory, but add qmd or hybrid retrieval once flat-file search becomes noisy at brain scale. Park's follow-up points to Membase as a hosted no-setup option; that is a fallback route, not Kevin's source of truth. Current GBrain source authority remains 0.42.56.0 at 646179047a8e4ad9c462d83ce9a67a50ba076de8. Source: X/@JoshuaIPark and self-thread, 2026-04-10/11; Source: Membase homepage, checked 2026-07-04; Source: GitHub garrytan/gbrain, 2026-07-04

Garry Tan SOUL.md completeness rule

New item in my SOUL md tonight - @garrytan | 2026-04-13 | 4,093 likes, 4,673 bookmarks

Takeaway: Promoted through Garry Tan and Builder Ethos. The reviewed SOUL.md screenshot is not merely vibes about intensity; it is an agent operating rule: when the permanent solve is within reach, avoid workarounds, search first, test before shipping, and return the finished product rather than a plan. Kevin's compiled version preserves the Builder Ethos lake/ocean scope guard so completeness does not become unbounded rewrite permission. Source: X/@garrytan, 2026-04-13; Source: local image review, 2026-07-03

Garry Tan - own and host your own memory

You should want to control and host your own memory. It's the one thing that you should be able to take to any platform. Watch for this to be a defining battle in the new browser war: the AI harness wars of 2027. - @garrytan | 2026-05-31 | 1,736 likes, 2,321 bookmarks

Takeaway: Memory portability as the next platform battleground; reinforces the markdown-as-memory, git-repo-as-brain thesis (memory survives any harness). Garry Tan, Agent Memory Patterns, The Brain-Agent Loop. Source: X/@garrytan, 2026-05-31

Learn Harness Engineering (course + awesome list)

This is the best site on the internet to learn harness engineering. - @_vmlops | 2026-05-17 | 4,028 likes, 7,367 bookmarks

Takeaway: Walking Labs project-based course plus the current 3,389-star awesome-harness-engineering list and 9,436-star course repo. Constrain behavior, maintain context, stop early victory, verify, observe. Dedicated page Learn Harness Engineering; concept Harness Engineering. Deep review keeps this as curriculum/reference, not an executable dependency. Source: X/@_vmlops, 2026-05-17

neo4j agent-memory - knowledge-graph memory layer

agent-memory by @neo4j is the best open-source repository for building a unified memory layer for AI agents via knowledge graphs. They did such an amazing job modeling short, long, and reasoning memory, their ontology, and extraction algorithms. - @pauliusztin_ | 2026-05-18 | 477 likes, 636 bookmarks

Takeaway: Promoted to Neo4j Agent Memory with artifact-level ontology evidence. The image shows user/fact/preference/document/chunk nodes, POLE+O nouns, custom objects, and typed relationships such as PART_OF, NEXT, and SUPERSEDED_BY; current source check records the repo at 355 stars, 82 forks, Python 0.5.0, and TypeScript 0.4.0. Maps to the wiki's compiled-truth + timeline + skills layers. Agent Memory Patterns. Source: X/@pauliusztin_ and local image/source review, 2026-07-03

implementation-notes.html as running agent memory

Implement . As you work maintain a running implementation-notes.html file that captures anything I should know about how the implementation [diverged from the spec]. - @trq212 | 2026-05-18 | 1,342 likes, 2,249 bookmarks

Takeaway: Deep review of the self-thread confirms the pattern is not "write a status log"; it is a permission boundary for ambiguity. The agent can make local implementation decisions, but must keep a running file of decisions, deviations, tradeoffs, and human-relevant context. Promoted into Harness Engineering and Agent Memory Patterns. Source: X/@trq212, 2026-05-18

ZeroEntropy - task-specific embedding/reranking models

A 6-person team is building task-specific AI models that are 4-8x faster than anything from OpenAI or Anthropic. 500K downloads on HuggingFace. This is what "make something people want" looks like in the model layer. - @garrytan | 2026-05-23 | 2,800 likes, 3,704 bookmarks

Takeaway: Deep-reviewed and promoted through ZeroEntropy plus Knowledge Brain System. Self-thread names ZeroEntropy as SOTA on GBrain embedding/reranking; current docs keep the model/API authority with zembed-1, zerank-2, SDKs, and endpoint details. Fast specialized rerankers are the retrieval half of RAG and agent memory, but the version/source facts belong on the dedicated tool page. Source: X/@garrytan thread, 2026-05-23; Source: ZeroEntropy docs/npm/PyPI/Hugging Face, 2026-07-03

Agent Skills & Configuration

Career-ops: AI job search system (open-sourced)

Bro created an AI job search system for Claude Code that scored 700+ job applications and actually got him a job. AND IT'S NOW OPEN-SOURCE. 14 skill modes. - @Hesamation | 2026-04-05 | 28,036 likes

Takeaway: Validates Writing and Content Skills skill pattern - skills can drive full pipelines, not just code assistance. Source: X/@Hesamation, 2026-04-05

Deep review: The current upstream is career-ops-v1.15.0 with 56K+ stars, and the local artifact shows a real TUI pipeline for scored offers, queue states, salaries, and application outcomes. Updated Writing and Content Skills, Santiago Fernández de Valderrama, and the executable career-ops skill with version context. Source: GitHub santifer/career-ops, 2026-06-30

andrej-karpathy-skills

Andrej Karpathy documented the exact ways LLMs fail at coding. Someone turned those observations into a single Claude config file. +3,741 stars this week. - @sharbel | 2026-04-11 | 4,677 likes

Takeaway: Andrej Karpathy's LLM Coding Skills is still reference-only, but the @sharbel screenshot is the cleanest artifact for the mapping: wrong assumptions -> think first, overengineering -> simple first, irrelevant edits -> surgical changes, no acceptance criteria -> goal-driven execution. Current source check found unchanged HEAD 2c606141936f1eeef17fa3043a72095b4765b9c2, no tags, no license field, 186,021 stars, and 19,088 forks; Kevin already owns these invariants through AGENTS, ethos, code-taste, iteration, and no-sus-code-doctor routes. Source: X/@sharbel, 2026-04-11; Source: GitHub README

Shaw's 8-subagent code quality prompt

The quality of your vibecoded slop is horrible. Use this prompt: "I want to clean up my codebase and improve code quality. This is a complex task, so we'll need 8 subagents." - @shawmakesmagic | 2026-04-15 | 5,948 likes

Takeaway: Subagent fan-out for code quality - each subagent owns one cleanup dimension. Parallels Cursor Orchestrate Skill. Source: X/@shawmakesmagic, 2026-04-15

"Caveman speech" token-saving skill

SOMEONE TURNED THE VIRAL "TEACH CLAUDE TO TALK LIKE A CAVEMAN TO SAVE TOKENS" STRATEGY INTO AN ACTUAL CLAUDE CODE SKILL - cuts ~75% of tokens while keeping full technical accuracy. - @om_patel5 | 2026-04-06 | 6,292 likes

Takeaway: Compression via output-style constraints - useful for long sessions, but keep as reference/candidate for now. Current source is JuliusBrussee/caveman, release v1.9.0, 81,611 stars, MIT; Brin returned caution for both repo and skill surfaces pending deep scan. See Caveman Token Skill. Source: X/@om_patel5, 2026-04-06; Source: GitHub and Brin API, 2026-07-03

65-line CLAUDE.md reduces mistake rate to 3-11%

40% of the code Claude writes for you is wasted. A 65-line markdown file fixes it. 120,000 developers have starred it. Mistake rate drop from 41% to 3-11%. - @dunik_7 | 2026-05-10 | 5,710 likes

Takeaway: Strong demand signal for compact CLAUDE.md/agent config, but the deep review did not verify the tweet's 120K-star repo or error-rate numbers. The local artifact is a Cal Rueb "Claude Code best practices" talk thumbnail/video. Preserve the durable lesson: recurring judgment belongs in harness files, skills, hooks, and checks, but version claims need primary sources. See Claude Code Compact Config Signal. Source: X/@dunik_7, 2026-05-10

Build systems that prompt themselves

Anthropic engineer: "You're not supposed to prompt Claude. You're supposed to build a system that prompts itself." Video breakdown of CLAUDE.md overhead and workflow mistakes. - @eng_khairallah1 | 2026-06-06 | 7,137 likes | 20,287 bookmarks

Takeaway: Harness > one-shot prompting; aligns with Harness Engineering and Kevin's AGENTS.md schema. Deep review promoted the concrete lesson into Claude Code Harness and Agent Looping: recurring judgement belongs in files, hooks, skills, and verification loops. Source: X/@eng_khairallah1 thread, 2026-06-06

Nick Spisak — twitter-bookmarks skill recipe

You should make a claude code skill called "twitter-bookmarks": read Chrome bookmarks, prioritize by recency, review against brand voice, output atomic implementation plan, human approve, then implement. - @NickSpisak_ | 2026-03-16 | 4,657 likes | 19,668 bookmarks

Takeaway: Describes the pipeline Kevin codified as Cloud, Data, and Service Skills + X Bookmarks to Wiki Pipeline. The reviewed self-thread adds the exact workflow: use a twitter-bookmarks skill to read Chrome-profile bookmarks, prioritize recent posts, compare against brand voice/mission, draft a comprehensive atomic todo plan, review it, approve it, then implement. Nick also generalizes the pattern for any repository: start Claude Code or Codex in plan mode, hand it the repo, research the codebase, interview it about benefits/features, ask for atomic implementation tasks, review, iterate, and only then approve implementation. Keep this as a skill/workflow reference, not a new local skill, because Kevin already owns the route through x-bookmark-absorb and the wiki pipeline. Source: X/@NickSpisak_ thread, 2026-03-16

Claude Code internal gates (reverse-engineered)

I reverse-engineered Claude Code's leaked source against billions of tokens of my own agent logs. Anthropic is aware of CC hallucination/laziness, and the fixes are gated to employees only. - @iamfakeguru | 2026-03-31 | 5,948 likes | 16,507 bookmarks

Takeaway: Source-leak artifact retained as provenance and a harness-audit checklist, but the employee-mode bypass claim is false. The canonical page is Claude Code Source Leak, and Claude Code Harness now frames the lesson as local verification, chunked reads, scoped searches, and stable harness files rather than a secret CLAUDE.md import. Source: X/@iamfakeguru thread, 2026-03-31

Also: Superpowers crosses from dev to marketer adoption (2,855 likes, @shannholmberg, 2026-03-14); deep review refreshed the current source to v6.0.3 / 242K+ GitHub stars and preserved the marketing workflow pattern: brainstorm, spec, task breakdown, parallel researcher/writer/editor/publisher agents, brand review, verify, publish. Source: compiled from X bookmarks; GitHub obra/superpowers, 2026-06-30

.claude folder as competitive moat

your .claude folder is what separates you from every other developer under the sun who uses Claude. it's your taste, rules, skills, and workflows. - @Hesamation | 2026-03-21 | 941 likes, 1,565 bookmarks

Takeaway: Promoted to Claude Code Memory Architecture (Reverse-Engineered). The reviewed image shows the useful mental model - managed policy, user, project, and local CLAUDE.md layers all combine into the session prompt - while current Claude Code docs add the caveat that instructions guide behavior but settings/hooks enforce hard controls. Source: X/@Hesamation and local artifact review, 2026-07-03

Hermes setup for "top 1% operators"

this is the Hermes setup top 1% operators are using to get rid of AI slop. - @EXM7777 | 2026-05-31 | 1,974 likes, 3,875 bookmarks

Takeaway: Promoted to The Eval Loop (Slop Is an Output Problem) as operator evidence, not just engagement signal. The reviewed diagram says slop is an output problem: prompts, larger models, memory, and context files are input-side guesses; the missing half is a 0-1 eval gate with a threshold and failure-feedback loop. Source: X/@EXM7777 and local artifact review, 2026-07-03

Agent Infrastructure & Platforms

Vercel open-sources cloud coding agent platform

Today we're open sourcing open-agents.dev, a reference platform for cloud coding agents. Stripe (Minions), Ramp (Inspect), Spotify (Honk), Block (Goose) are building their own "AI software factories." - @rauchg | 2026-04-14 | 4,277 likes

Takeaway: Sandboxes + Workflow SDK + AI Gateway reference stack. Open Agents.dev is versioned by source commit, not npm: HEAD cf865e94de7729751c747171785b6ce57e7b178c, no release tags, 5,693 stars, Brin caution/pending. Source: X/@rauchg, 2026-04-14

Bud - AI Human Emulator

Introducing Bud. The first AI Human Emulator. Full computer with storage, compute, memory. Browser, SMS, Telegram. Learn custom skills. Work fully autonomously. - @budapp | 2026-04-21 | 4,026 likes

Takeaway: Full-computer agent with persistent skills, browser fallback, scheduled automations, and web/text/Telegram reachability - agent-as-contact plus agent-as-employee. Deep review refreshed Bud - AI Human Emulator against current Bud docs and treats launch SMS numbers as volatile product entry points rather than stable facts. Source: X/@budapp, 2026-04-21; Source: bud.app, 2026-07-01

Higgsfield Supercomputer

The first ever cloud-native, self-learning AI agent for end-to-end task execution. 40+ built-in tools. Three layers of memory. Powered by enhanced Hermes Agent. - @higgsfield | 2026-05-13 | 3,371 likes

Takeaway: Deep review keeps Higgsfield Supercomputer as a creative-agent product reference, not a local dependency. The durable idea is the visible agent-product grammar: role agents/"AI employees", connectors, slash-triggered skills, persistent files, scheduled tasks, and memory as first-class user-facing primitives. That belongs in Agent Harness Capsules and Brain Capsules as product shape evidence for persistent creative agents. Source: X/@higgsfield, 2026-05-13; Source: Higgsfield Supercomputer product page, 2026-06-30

Vercel Workflows GA

Vercel Workflows is GA. Your code is the orchestrator. Ship agents, backends, or any long-running process without managing queues, retries, or workers. - @vercel | 2026-04-16 | 447 likes

Takeaway: Durable execution for agents goes GA on Vercel - validates workflow skill. Source: X/@vercel, 2026-04-16

Also: Vercel Sandbox persistence - the March beta introduced durable sandbox names with session-level compute: stopping a session snapshots the filesystem, resuming starts a new VM from that snapshot, and state storage is free while compute is only billed when running. By July 2026 the npm packages had moved to @vercel/sandbox@2.4.0 and sandbox@3.4.0; the beta dist-tags remained available for the original changelog path. See Persistent Sandbox Thesis and Agent Sandboxes. Source: Vercel changelog, 2026-03-26; Source: Vercel Sandbox docs, 2026-07-04; Source: npm registry, 2026-07-04

Sandboxes are becoming the OS for agents

Sandboxes are becoming the OS for agents. This is because agents will be more persistent and stateful. I go deep in my Substack: isolation primitives, performance, portability, architecture, build vs buy, market dynamics. - @crystalxtang | 2026-05-27 | 272 likes, 299 bookmarks

Takeaway: Substack analysis of the sandbox layer: three workload shapes, microVM vs container isolation, filesystem-as-memory, harness inside vs outside, and the Modal/e2b/Daytona build-vs-buy split. Dedicated page Agent Sandboxes now also preserves the local provider-map image: E2B/Vercel/Cloudflare cluster around stateless lightweight execution; Daytona/Runloop/declaw/CodeSandbox/Hopx around stateful lightweight environments; Modal/Northflank/Tensorlake around stateless compute-heavy workloads. See Open Agents.dev, Browser Testing Skills. Source: X/@crystalxtang, 2026-05-27; Source: local image artifact review, 2026-07-02

Browser Automation

Browserbase open-source: agents learn sites, save skills

Do you understand what Browserbase just open-sourced??? An agent that learns any website once, then does the job 10x cheaper forever. - @DeRonin_ | 2026-05-08 | 2,061 likes

Takeaway: Community validation of Autobrowse (Browserbase Skills) - "learn once, run forever" pattern resonates. Source: X/@DeRonin_, 2026-05-08

Lightpanda - headless browser for agents

Someone built a headless browser from scratch that runs 11x faster and uses 9x less memory. It's called Lightpanda. Every AI agent doing web automation is running Chrome under the hood. - @ihtesham2005 | 2026-03-15 | 1,350 likes

Takeaway: Lightpanda is a browser-efficiency candidate for fast, low-memory agent reading, not the default scraping engine. Treat it as speed/token infrastructure for cooperative pages. Deep review on 2026-07-02 refreshed the source snapshot to release 0.3.4, HEAD 60aacfa2b064877c3eecddf98a5dd61048c1bdd8, 31.5K stars, and 1.4K forks; the reviewed image's 18K-star/11x/9x claim is historical launch evidence. For protected targets, the later leftcurve thread correctly changes the route: anti-bot success depends on TLS and browser-fingerprint realism, not only DOM speed. Use Web Scraping Stealth (TLS Fingerprinting & Anti-Bot Evasion) and Camofox Browser before recommending Lightpanda for adversarial scraping. Source: X/@ihtesham2005, 2026-03-15; Source: GitHub lightpanda-io/browser, 2026-07-02; Source: X/@leftcurvedev_, 2026-04-09

Camofox Browser for stealth scraping

As someone who scraped for a living for years, anyone recommending Lightpanda for scraping does not understand the hard part: TLS fingerprinting and browser realism. - @leftcurvedev_ | 2026-04-09 | 2,287 likes, 3,323 bookmarks

Takeaway: Added Camofox Browser as the explicit stealth-browser route. The artifact is not a generic browser demo: it is an anti-detection browser server for AI agents, built on Camoufox with Playwright compatibility, stable element references, Docker, and token-efficient accessibility snapshots. Version check on 2026-06-30 found GitHub release v1.11.2, npm @askjo/camofox-browser@1.11.2, and npm namespace collisions around unscoped camofox-browser; use jo-inc/camofox-browser as the authority. Source: X/@leftcurvedev_, 2026-04-09; Source: jo-inc/camofox-browser, 2026-06-30

Browser Use Box - 24/7 persistent agent

Introducing Browser Use Box (bux). Your 24/7 personal agent box. Real Chrome with persistent logins. Telegram baked in. Runs while you sleep. - @larsencc | 2026-04-26 | 852 likes

Takeaway: Always-on browser agent with persistent auth - the "agent employee" pattern extends to browser tasks. Deep review refreshed Browser Use Box (Bux) from the public browser-use/bux repo, README, GitHub API, local image, and local video contact sheet. Current route: source-owned Bux page, not hub-only launch note. Source: X/@larsencc, 2026-04-26; Source: GitHub browser-use/bux, 2026-07-02

browser-to-api - reverse-engineer any website into an API

Turn any website into an API with /browser-to-api. Analyzes network activity, CDP logs, and website behavior to generate a custom OpenAPI spec. - @derekmeegan | 2026-05-13 | 2,428 likes

Takeaway: CDP-level site analysis can become a generated API hypothesis, but the reviewed Browserbase source makes this a gated workflow: capture with browser-trace, generate with Browser to API, inspect/redact reports, and only then build a durable tool. Complements Browser Testing Skills for structured extraction without replacing official APIs or ToS review. Source: X/@derekmeegan, 2026-05-13

Cost Optimization & Token Efficiency

lean-ctx: 60-95% token cost reduction

Reduces Claude Code and Cursor token costs by 60-95% - @tom_doerr | 2026-05-08 | 324 likes

Takeaway: See Lean Ctx; deep review refreshed current GitHub/npm v3.8.17, 81 MCP tools, 30+ agents, and screenshot drift from the earlier v3.4.5 / 49-tool launch image. Source: X/@tom_doerr, 2026-05-08

Polish language → 60% more tokens for Anthropic

Text written in Polish translates to 60% more tokens for Anthropic models. You can get your app 60% faster and cheaper simply by switching languages. - @_overment | 2026-03-07 | 2,305 likes

Takeaway: This is Language Orchestration / Tokenizer Agent Interaction evidence, not a universal "switch languages" rule. The reviewed screenshot shows en.txt: 36,637 vs pl.txt: 59,442 tokens (+62.25%), while the author's self-thread narrows it to "up to," recommends tokenizer/provider estimation endpoints, and warns that quality/performance claims are task/model dependent. Source: X/@_overment thread, 2026-03-07

Agent UX & Interfaces

Stay in the loop — Suzanne's teaching prompt

been asking others at Anthropic how they stay in the loop with Claude and fully understand the work being done — this is one of my favorites from Suzanne. - @trq212 | 2026-06-01 | 10,413 likes | 17,809 bookmarks

Takeaway: Anthropic-internal pattern for human comprehension, not rubber-stamp approval. Deep review inspected the gist and screenshot, then chose no new duplicate skill: it folds into existing Staying in the Loop with Agents, Security and Review Skills, and Security and Review Skills. Pairs with voice mode. Source: X/@trq212 thread, 2026-06-01

Thariq X article (link-only bookmark)

@trq212 | 2026-06-02 | 10,221 likes | 23,358 bookmarks

Takeaway: High-bookmark link-only post to an X Article; destination blocked at enrichment time — hub entry only until article content is recoverable. Source: X/@trq212, 2026-06-02

Lance Martin X article (link-only bookmark)

@RLanceMartin | 2026-06-09 | 5,960 likes | 15,729 bookmarks

Takeaway: Source review could not recover the X Article body: the resolved long-form article still blocks anonymous fetch. Search indexes identify the article as Lance Martin's loop-engineering piece tied to Anthropic's "Building Effective Agents", but the exact body remains unverified. Keep the durable synthesis in Agent Looping and do not promote article-specific claims until authenticated re-enrichment recovers the source text. Source: X/@RLanceMartin, 2026-06-09; Source: source review, 2026-07-04

Farza - voice-first agent UX

Been working on a new UX for agents. It will understand your workflow as you use your computer and proactively nudge you when it can be helpful. Agents are super capable beings trapped in terminals and chat interfaces. - @FarzaTV | 2026-05-16 | 1,053 likes

Takeaway: Voice + ambient awareness as agent UX - escape the chat/terminal paradigm. Deep review keeps this as Farza / Clicky / Agent Product Surface evidence: the video moves from talking-head framing into a Mac desktop overlay with Home/Agents navigation, hold-to-control and question-to-talk hints, cursor controls, app integrations, and an "Unlock Clicky" action. The product-shape lesson is a companion control panel beside the desktop, not a transcript-only chat. Source: X/@FarzaTV, 2026-05-16; Source: local video contact sheet, 2026-07-02

Clicky open-sourced (Farza's agent UX)

Hey, I'm open-sourcing Clicky. Go forth into the wild and build the future of education and the future of AI interfaces. - @FarzaTV | 2026-04-08 | 5,852 likes

Takeaway: Clicky now has a focused tool page. The repo is a Swift macOS menu-bar app with ScreenCaptureKit context, AssemblyAI transcription, Claude streaming, ElevenLabs TTS, coordinate pointing, and a Cloudflare Worker secret proxy. Current source authority: MIT, HEAD a80fa80721a8aebe51a170a7780705024ebc6e46, no GitHub release, 6,946 stars, 1,342 forks. No local executable skill yet because this is app prior art, not a repeated Kevin workflow. Source: X/@FarzaTV, 2026-04-08; Source: GitHub farzaa/clicky, 2026-06-30

Codex mobile + Tailscale

Codex mobile app can now manage all my devices through tailscale. Incredible. OpenAI won. - @redtachyon | 2026-05-15 | 3,315 likes

Takeaway: Promoted through Tailnet Agent Mesh - Reaching Your Agents from Anywhere as field evidence for mobile agent control over a private device mesh. The reviewed screenshot shows Codex mobile listing multiple online hosts with terminal/device icons; OpenAI's current docs match the architecture: phone as control surface, host machine as execution plane, and VPN/mesh networking instead of public app-server exposure. Source: X/@redtachyon, 2026-05-15; Source: OpenAI Codex mobile, 2026-07-03; Source: OpenAI remote connections, 2026-07-03

Agent Criticism & Counter-Signal

Mitchell Hashimoto on AI psychosis

I strongly believe there are entire companies right now under heavy AI psychosis and its impossible to have rational conversations about it with them. - @mitchellh | 2026-05-15 | 15,251 likes, 4,831 bookmarks

Takeaway: Source-reviewed and already carried by Mitchell Hashimoto plus Builder Ethos. The durable nuance is not anti-agent posture: infrastructure can recover through bounded MTTR, but libraries, desktop, and mobile are limited by user upgrade cycles; agent-authored code still needs 100% review in whole-system context. Source: X/@mitchellh and self-thread, 2026-05-15

Vibecoding reality check

Vibecoder asks claude code to build a chat app, gets a working prototype in 20 minutes, immediately tweets "just killed slack and discord"… brother you don't even know what a distributed system is. - @paoloanzn | 2026-03-13 | 4,924 likes

Takeaway: Dunning-Kruger in vibecoding - building a prototype ≠ building infrastructure. See Vibecoding Critique. Source: X/@paoloanzn, 2026-03-13

Agent Tools & Utilities

React Doctor v2

React Doctor v2 is here. Your agent writes bad React code, this catches it. Works with Next.js, Vite, React Native. npx react-doctor@latest - @aidenybai | 2026-05-08 | 5,161 likes

Takeaway: Agent code quality tooling - Security and Review Skills validated at scale. Source: X/@aidenybai, 2026-05-08

MarkItDown - Microsoft's anything-to-markdown

MICROSOFT BUILT A TOOL THAT CONVERTS LITERALLY ANYTHING INTO CLEAN MARKDOWN FOR YOUR LLM. PDFs, Word docs, Excel, PowerPoint, audio, YouTube URLs. - @_vmlops | 2026-04-08 | 4,957 likes

Takeaway: Universal file-to-markdown pipeline - eliminates custom parsers for agent ingestion. Deep review fixed the canonical source to microsoft/markitdown (0.1.6) and added the key security caveat: only convert trusted/local inputs or narrow the converter method. See MarkItDown - Universal File-to-Markdown Converter. Source: X/@_vmlops, 2026-04-08; Source: X/@mdancho84, 2026-05-27

HyperFrames - agent-native video (HeyGen)

We built our launch video in Claude Code using HyperFrames. Open source, agent-native framework. HTML to MP4. - @HeyGen | 2026-04-16 | 8,226 likes

Takeaway: HyperFrames is the agent-native HTML-to-MP4 pipeline, installable as a skill. Deep launch-video review shows a speaker, agent/HTML surface, generated motion examples, a prompt surface, and rendered end card; the current version check moved the tool to hyperframes@0.7.24. Source: X/@HeyGen, 2026-04-16; Source: npm registry + local video artifact review, 2026-07-01

HyperFrames in Grok

Takeaway: HeyGen's later Grok bookmark upgrades HyperFrames from coding-agent-only to a chat-surface connector. Grok's connector directory now lists HyperFrames by HeyGen for programmatic HTML video, while HeyGen's broader avatar/video tools remain a custom MCP connector. Use the Grok path for fast ideation; use the local skill/CLI path for repo-controlled deterministic content pipelines. Source: X/@HeyGen, 2026-06-17; Source: HeyGen Grok docs, 2026-06-30

ml-intern - automated ML research agent

Introducing ml-intern, the agent that automated the post-training team @huggingface. Researches papers, goes through citations, implements ideas in GPU. - @akseljoonas | 2026-04-21 | 4,663 likes

Takeaway: Research agent that reads papers, follows citations, implements on GPU, and can run through HF/local model and sandbox routes. Deep review updated ML Intern - Automated ML Research Agent with repo source, 0.1.0 pyproject version, no public tags/releases, and a caution that launch benchmark media is not a reproducible benchmark by itself. Source: X/@akseljoonas, 2026-04-21; Source: GitHub huggingface/ml-intern, 2026-07-02

Also notable: Shaders MCP (943 likes, @npm_i_shaders) - domain-specific MCP for shader discovery, installation, prop wiring, and SVG-to-SDF generation from Shaders Pro. Deep review routed the launch demo through Shaders MCP as an agent-native asset library pattern. Abstract Chain-of-Thought (1,072 likes, @KeshavRamji) - compressed reasoning via learned abstract tokens. spawn-agent (407 likes, @aidenybai) - universal API to invoke any coding agent. Source: compiled from X bookmarks; Shaders MCP docs, 2026-07-01

UI as a state machine for testing

representing ui as a state machine makes it easier to manually and automatically test. - @JohnPhamous | 2026-05-26 | 781 likes, 752 bookmarks (demo video)

Takeaway: Promoted to React Style as a testing rule. The local artifact shows an MFA modal harness with named scenarios, avatar state controls, error injection, production-flow coverage notes, status cards, and an event log. Same principle as the Frontend and Design Skills skill applied to frontend testability: enumerate the user-touchable state machine before trusting one happy-path render. Source: X/@JohnPhamous, 2026-05-26; Source: local artifact review, 2026-07-03

SEO & AI Search

GEO-SEO tooling gaining community traction (2,354 likes, @tom_doerr, 2026-03-09) - validates Security and Review Skills skill. Claude Code + SEO APIs (.env with Keywords Everywhere, DataForSEO, GSC, CMS access) = automated optimization pipeline (1,258 likes, @codyschneider, 2026-03-02). Deep review of @alexgroberman's John Mueller row promoted an Writing and Content Skills measurement loop: keep Search Console clicks/impressions, track AI answer mentions/citations, inspect AI Overview citation columns where tools expose them, and treat AI search as audience-behavior measurement rather than a magic separate SEO discipline. Source: X/@alexgroberman, 2026-03-12; Source: local artifact review, 2026-07-03

Multi-Agent Systems

Matt Epstein - 21-Agent Viral Launch System

We do not use Claude as a writer. We use Claude as an operating system for launches. Inside Claude Code, we run a group of 21 specialized agents. Each agent has one job. - @mattepstein | 2026-05-14 | 1,036 likes, 4,758 bookmarks

Takeaway: Full launch pipeline as multi-agent orchestration: research -> BOLD CLAIM positioning -> hook writing with adversarial critique -> weapons check (invention novelty + copy intensity) -> Mom Test for mass-market comprehension. The linked X Article is currently blocked to anonymous fetches, but the enriched self-thread and artifact reply preserve the Mom Test wording plus the four-question launch brief. 26/30 launches went viral. See Launch Agent System (Matt Epstein). Source: X/@mattepstein and enriched row, reviewed 2026-07-04

Hermes Agent - Self-Evolving Agent Framework (90K+ stars)

Hermes Agent crossed 90,000 GitHub stars in two months. Three-tier memory, self-evolving skills, GEPA offline optimization, and multi-profile agents. - @akshay_pachaar | 2026-05-13 | 4,857 likes, 18,444 bookmarks

Takeaway: Validates Kevin's wiki architecture, but the current source split matters: Hermes Agent (Nous Research) owns memory/profile/agent-stack details, while GEPA - Genetic-Pareto Skill Evolution owns the optimizer. Deep review checked Akshay's article/thread, the YouTube crash-course transcript, arXiv v2, GitHub, and PyPI. GEPA is now a concrete MIT Python package (gepa@0.1.1, tag v0.1.1, HEAD 92dadfffbe98c8ecf508179a1cab09c1bb85cd32) that optimizes textual system components from traces using reflection, mutation, and Pareto selection. Source: X/@akshay_pachaar, 2026-05-13; Source: arXiv:2507.19457; Source: GitHub gepa-ai/gepa

Claude Code /goal - Autonomous Execution with Model-Verified Exit

/goal is the most time-saving feature in all of AI right now. Set a completion condition. Claude works autonomously, turn after turn, until that condition is met. - @NickSpisak_ | 2026-05-14 | ~4,900 likes, ~19K bookmarks from earlier import

Takeaway: Model-verified exit conditions for autonomous agent runs. Haiku as fast evaluator. Requires specific, measurable goals + CLAUDE.md + hooks + auto mode. See Claude Code /goal Command. Source audit: the previously recorded status ID belongs to @mattepstein, so the Nick Spisak source URL is unresolved until recovered from another export. Source: local wiki import, reviewed 2026-07-04

Solo dev's 38-agent Claude Code hackathon stack

A solo dev just won the Anthropic hackathon, shipped a working product in 8 hours with Claude Code, and walked away with $15,000. Then he open-sourced the entire stack. 153,000 stars on GitHub. 38 specialized agents. - @PrajwalTomar_ | 2026-05-23 | 4,322 likes, 7,792 bookmarks

Takeaway: Deep artifact review resolves this as Everything Claude Code - Agent Harness Performance System evidence: the screenshot names 38 agents, 156 skills, 72 commands, and 1,282 security tests, plus a selective install path. Use it as an adoption/source artifact while the canonical version authority stays the current affaan-m/ECC repo and npm release snapshot. Source: X/@PrajwalTomar_, 2026-05-23

Agent Marketing & Content

Austin Lau - Anthropic's one-person marketing operation

Anthropic ran its marketing with basically one person. Austin Lau, a non-technical growth lead, was running paid search, paid social, email, and SEO solo. Export ad CSVs into Claude Code → AI flags underperforming ads → agents generate replacements. - @cgtwts | 2026-03-14 | 6,567 likes

Takeaway: One person + agents = full marketing team - Services-as-Software in action at Anthropic itself. Deep video review preserves this as Austin Lau person/workflow evidence, not a new tool: Claude Code sits over CSV export, underperforming-ad analysis, replacement generation, email, SEO, paid search, and paid social. Source: X/@cgtwts, 2026-03-14

Skill graph for content teams

I run 10 social media accounts and don't write a single post manually. The secret: a skill graph. 30+ markdown files wired together that turned my AI agent into a full content team. - @DeRonin_ | 2026-03-14 | 2,847 likes

Takeaway: Skill graph as content team - markdown files → wired agents = 10 accounts automated. Source: X/@DeRonin_, 2026-03-14

last30days v3 - 20K+ stars

v3 of @slashlast30days is here. 20,000+ stars on GitHub. An AI agent-led search engine scored by upvotes, likes, and real money - not editors. Reddit, X, YouTube are now FREE. - @mvanhorn | 2026-04-09 | 1,083 likes

Takeaway: Security and Review Skills hits 20K+ stars - engagement-scored search validated by community. Source: X/@mvanhorn, 2026-04-09

Design & Agent UI Quality

Awesome Design MD - 30+ company design systems in one file

Someone reverse-engineered the design systems of Apple, Spotify, Airbnb, and 30+ companies. Packed each one into a single file. Drop one file into your project. Your AI agent builds UI that looks like Spotify. - @heynavtoor | 2026-04-04 | 8,011 likes

Takeaway: Design system -> single .md file -> agent-consumable. Deep review corrected the source repo to VoltAgent/awesome-design-md; Google design.md is the spec node, not this collection. See Awesome Design MD. Source: X/@heynavtoor, 2026-04-04

Custom ESLint rules as anti-slop enforcement

YOU GUYS NEED TO PUT YOUR AGENTS ON CUSTOM ESLINT RULES ASAP. IT IS THE BEST WAY TO GUARANTEE ANTI-SLOP IN YOUR CODEBASE BY MAKING IT IMPOSSIBLE TO DO SLOP PATTERNS. - @KingBootoshi | 2026-04-06 | 730 likes

Takeaway: Deep review strengthened Lint-Enforced Agent Guardrails: the local artifact shows three concrete custom rules (no-unscoped-service-test, require-error-code-assertion, no-mock-echo) and the self-thread generalizes the pattern to pylint, clippy, and clang-tidy for non-TS codebases. Source: X/@KingBootoshi and local artifact review, 2026-07-03

Higgsfield MCP - motion design in Claude Code

Motion design with Claude Code + Higgsfield MCP: 1. Pull references with Pinterest API. 2. Storyboard 6 scenes via GPT Images 2.0. 3. Drop the board into Seedance 2.0. Motion Design, solved. - @higgsfield | 2026-05-18 | 2,688 likes, 2,739 bookmarks (demo video)

Takeaway: Full motion-design pipeline (Pinterest references -> GPT Images 2.0 storyboard -> Seedance 2.0 video) orchestrated from Claude Code via Higgsfield MCP. Deep review confirmed the local video as an end-to-end Claude Code/Higgsfield workflow, not just a generated-video reel, and refreshed the current source snapshot to @higgsfield/cli@1.1.1 plus Claude connector https://mcp.higgsfield.ai. Source: X/@higgsfield, 2026-05-18; Source: Higgsfield docs/npm/GitHub, 2026-07-02

Higgsfield MCP - "marketing agency in your laptop"

18-minute breakdown of Claude + Higgsfield MCP: 1. Research across TT, IG, YT. 2. Production calendar by the agent. 3. UGC, motion design, carousels. 4. Approval gate before spend. 5. Integrate Meta Ads MCP. - @higgsfield_ai | 2026-05-18 | 1,815 likes, 2,540 bookmarks (demo video)

Takeaway: End-to-end content/ad agency workflow with a human approval gate before ad spend, chaining Higgsfield MCP and Meta Ads MCP. Deep review confirmed agent UI, production plan/spec text, UGC assets, carousel/video creative, and "STAGE 4 META ADS" handoff frames in the local video contact sheet. Source: X/@higgsfield_ai, 2026-05-18; Source: local video review, 2026-07-02

Security, Red-Teaming & Supply Chain

Cloudflare tests Anthropic's Mythos against 50 of its own repos

Cloudflare's security team spent the last few weeks testing Anthropic's Mythos against fifty of our own repositories. What we learned about offensive AI, why faster patching is the wrong reaction, and what the architecture around vulnerabilities has to look like next. - @Cloudflare | 2026-05-18 | 3,964 likes, 4,805 bookmarks

Takeaway: First-party validation of Project Glasswing's Mythos Preview at vendor scale; the lesson is architectural (assume offensive AI) rather than "patch faster." Source: X/@Cloudflare, 2026-05-18

Cloudflare's vulnerability-discovery harness pipeline

Cloudflare on their vulnerability discovery harness: Recon (return an architecture doc), Hunt (~50 agents look for bugs concurrently), Validate (independent agents try to disprove findings), Gapfill, Dedup. - @eugeneyan | 2026-05-18 | 303 likes, 344 bookmarks

Takeaway: Eugene Yan's distilled stages of the Project Glasswing harness; narrow-scoped parallel agents plus adversarial validation beat one exhaustive agent. Full 8-stage table in Harness Engineering. Source: X/@eugeneyan, 2026-05-18

"Red Team Mode" - 7 adversarial business prompts

Claude has a feature called Red Team Mode. You can use it to attack your own business the way a competitor, investor, or angry customer would, and fix the weak spots before they become real problems. Here are 7 prompts. - @free_ai_guides | 2026-05-27 | 2,665 likes, 4,918 bookmarks

Takeaway: Not an actual Claude feature but a reusable seven-prompt adversarial self-audit (competitor war game, churn sim, investor teardown, SPOF scan, market shift, pricing pressure, attack-surface report). Dedicated page Adversarial Self-Audit Prompts. Source: X/@free_ai_guides thread, 2026-05-27

npm supply-chain hardening for the AI-attack era

Every week brings a new neural-network-assisted breach. Make time to check: are all dependencies really needed; use only npm ci in CI; host your own dependency registry; use npm install --min-release-age=7. - @tshemsedinov | 2026-05-18 | 89 likes, 105 bookmarks

Takeaway: Promoted through npm Supply Chain Security and Brin - Agent Context Security. The durable guidance is dependency minimization, lockfile-only CI installs, private registries where governance matters, and release-age quarantine only when the runner's package manager actually enforces the config. Source: X/@tshemsedinov, 2026-05-18

Budget AI cybersecurity with Cursor + GPT 5.5

So here's how you have AI cyber sec setup on a budget: Use @cursor_ai with @OpenAI (GPT 5.5) to understand the system and explore possibilities across the architecture, and then try everything till you find one. btw OpenAI bypassed Claude's Opus 4.6. - @sathuashrith | 2026-05-27 | 34 likes, 68 bookmarks

Takeaway: Deep review reframed this as an Agent Security Model signal, not a recommendation to "try until you find one." The reviewed images show the useful boundary: safe refusal to retrieve leaked files, paired with a reportable finding that tool traces/WebSocket logs exposed paths, tool capabilities, and potentially .env content. Route future rows like this to tool-trace redaction, observability, and safe-disclosure policy. Source: X/@sathuashrith, 2026-05-27; Source: local artifact review, 2026-07-03

Adjacent / Cross-Domain Signal

Bookmarks in this batch that sit outside the AI/agents core but were captured alongside it:

• Product growth minimum requirements - Jeff Weinstein on why products stall: you need a frequent value moment, real pain, and a differentiated wedge against crowded substitutes. Promoted into Devtool Growth Playbook as the pre-channel gate before distribution tactics. (792 likes, 932 bookmarks, @jeff_weinstein, 2026-05-27)

- **Supabase getUser() -> getClaims()** - promoted into Security and Review Skills and the Supabase Auth skill. `getClaims()` can avoid repeated Auth-server requests through cached JWKS verification on asymmetric signing keys; `getUser()` always calls the Auth server, but remains the stronger pattern for freshness-sensitive authorization. (514 likes, 290 bookmarks, [@supabase](https://x.com/supabase/status/2058919774051053834), 2026-05-25) - **Cloudflare for Startups credits** - source-reviewed into Cloudflare: the relaunched program offers up to $350k in one-year credits across $10k / $100k / $350k tiers, covers Workers, Durable Objects, Workflows, Workers AI, KV, D1, Queues, Vectorize, R2, Pages, Images/Stream, Cache Reserve, and Argo Smart Routing, and excludes AI Gateway credits for now. (1,923 likes, 1,816 bookmarks, [@kristianfreeman](https://x.com/kristianfreeman/status/2059188629780545973), 2026-05-26; [Cloudflare Startups, 2026-07-04](https://www.cloudflare.com/startups/)) - **FMHY (Free Media Heck Yeah)** - viral thread on the volunteer-run free-tools/media index; author clarifies it is an index that hosts nothing. Deep review keeps it as an open-web directory/curation case study, not an active agent route for copyrighted-media sourcing. (18,921 likes, 33,203 bookmarks, [@ihtesham2005](https://x.com/ihtesham2005/status/2058221813797446075), 2026-05-23) - **Web Share API JS tip** - promoted to Web Share API with current MDN constraints: secure context, transient user activation, `web-share` Permissions Policy, cancellation handling, and `navigator.canShare()` for files. The durable product move is replacing stale social-button grids with one native share sheet when browser support and privacy constraints fit. (829 likes, 617 bookmarks, [@denicmarko](https://x.com/denicmarko/status/2055589633795461181), 2026-05-16) - **Santander AI Open Source** - promoted from the screenshot bookmark into a focused tool page. Live source review updates the launch count from 11 repos to 14 public repositories and keeps the shelf as a financial-services AI governance/source-reference signal, not a local skill import. (1,184 likes, 1,779 bookmarks, [@sytaylor](https://x.com/sytaylor/status/2068684111183495368), 2026-06-21; [GitHub org, checked 2026-07-04](https://github.com/SantanderAI)) - **Unresolved captures** - [@ankur_d](https://x.com/ankur_d/status/2056580180211093571) (326 bookmarks), [@subahwadhwani](https://x.com/subahwadhwani/status/2056384331762090026) (1,960 bookmarks, content-engine promo thread; rechecked 2026-07-04 and still only a logged-out X Article shell), and [@akshay_pachaar](https://x.com/akshay_pachaar/status/2054915602171723992) (504 bookmarks) all linked X long-form articles that resolved only to a login wall; exact-id web search on 2026-07-04 did not find a public mirror, so no extractable content can be promoted. Source: compiled from enrichment, 2026-06-01; source review, 2026-07-04

Newest 55 — Jun 2026 sync (2026-06-13)

Polished pass on the 55 bookmarks added in the latest ft sync. Dedicated pages: Claude Fable 5, Shopify Quick, Roboflow Supervision, Frontend and Design Skills, Contentbit, Text-to-Lottie, Extend UI, Agent Looping.

Claude Fable 5 launch

Introducing Claude Fable 5: a Mythos-class model that we've made safe for general use. Its capabilities exceed those of any model we've ever made generally available. - @claudeai | 2026-06-09 | 104,049 likes, 21,895 bookmarks

Takeaway: General Mythos-class release with topic fallbacks to Opus 4.8; restricted Claude Mythos 5 for Glasswing partners. Full page Claude Fable 5; security context Project Glasswing. Source: X/@claudeai thread, 2026-06-09

Fable 5 for self-improving agent systems

"Fable 5 is our best model for running self-improving agent systems. Add /loops, dynamic workflows, dreaming and you are unstoppable" — Anthropic Managed Agents team (13-min build-from-scratch demo). - @0xCodez | 2026-06-11 | 1,951 likes, 5,001 bookmarks

Takeaway: Positions Fable 5 as the model layer for Agent Looping + The Eval Loop (Slop Is an Output Problem) harnesses. Demo video on bookmark. Source: X/@0xCodez, 2026-06-11

Deep review upgrades this from a model-hype line into a platform-loop artifact: the demo frames longer task horizons as an access-boundary problem, then shows Claude Managed Agents as agent definition, environment configuration, session execution, and event handling. Route the reusable lesson to Agent Looping: autonomy needs event topology, credentials boundaries, memory, and verifiable outcomes. Source: local video review, 2026-07-03

Claude Code Workflows

Workflows are the biggest upgrade to Claude Code's capabilities since skills and subagents. - @trq212 | 2026-06-08 | 4,690 likes

Takeaway: Anthropic engineer framing for durable multi-step agent flows — complements Claude Code Harness skills/subagents. Source: X/@trq212, 2026-06-08

Staying in the loop with Claude Code

been asking others at Anthropic how they stay in the loop with Claude and fully understand the work being done — this is what they do - @trq212 | 2026-06-07 | 10,413 likes

Takeaway: Internal Anthropic patterns for human oversight during autonomous work. See Staying in the Loop with Agents. Source: X/@trq212, 2026-06-08

Fable edited its own launch video

Lots of people asked how I used Fable to edit its own launch video — TLDR it wrote a lot of tooling - @trq212 | 2026-06-10 | 9,154 likes

Takeaway: Meta demo of Claude Fable 5 on long-horizon video tooling, now deep-reviewed as a concrete media pipeline: transcription, subagent take selection, JSON edit decision list, ffmpeg first cut, generated LUT options, Remotion JSX graphics, and transcript-timed cue sheets. The follow-up caveat is useful: the crew called the result "okay", so the durable lesson is code-shaped media production plus human taste loops, not autonomous art direction. Source: X/@trq212, 2026-06-10; Source: video-editing deck, 2026-06-10

shadcn /improve — smart model plans, cheap model executes

Introducing /improve: use your most capable model to audit your codebase and write plans for cheaper models to execute later. - @shadcn | 2026-06-10 | 6,151 likes, 8,716 bookmarks

Takeaway: shadcn/improve is the concrete Borrowed Intelligence (Drain It into Plans) workflow: run the strongest model for audit, discovery, plan writing, review, backlog reconciliation, or GitHub issue publication, then hand bounded plans to cheaper executors. The later follow-up makes the model-risk argument explicit: if frontier models can be pulled or repriced, durable plans are how temporary intelligence becomes inventory. Source: X/@shadcn thread and local image review, 2026-06-10; Source: X/@shadcn, 2026-06-13

Shopify Quick — internal zero-config web platform

Everyone's talking about AI-generated HTML. But have you tried giving your sites a zero-config API for saving data, file storage, AI, websockets? We call it Quick — 50k+ internal sites on one VM. - @pushmatrix | 2026-06-10 | 2,902 likes, 5,712 bookmarks

Takeaway: High-trust internal prototype platform (not public). Page Shopify Quick. Source: X/@pushmatrix thread, 2026-06-10

Loops beat prompts (Anthropic engineer framing)

"You're not supposed to prompt Claude. You're supposed to build a system that prompts itself." - @eng_khairallah1 | 2026-06-06 | 7,137 likes

Takeaway: Same thesis as Agent Looping and @0xCodez / @divaagurlxw harness-engineering posts in this batch. Source: X/@eng_khairallah1, 2026-06-06

Claude Code creator on loops

"I don't prompt Claude anymore. I write loops — and the loops do the work." - @0xCodez | 2026-06-05 | 3,982 likes

Takeaway: Agent Looping social proof from Claude Code creator discourse. Source: X/@0xCodez, 2026-06-05

Harness engineering > prompt engineering

As an AI Engineer. Please learn harness engineering, not just prompt engineering; context engineering, not just RAG. - @divaagurlxw | 2026-06-05 | 4,277 likes

Takeaway: Duplicate/cross-batch reference for the reviewed row above; canonical update now lives in Harness Engineering. Source: X/@divaagurlxw, 2026-06-04

Host your own memory

You should want to control and host your own memory — it's the one thing that you should be able to take to any platform - @garrytan | 2026-05-31 | 1,736 likes

Takeaway: Garry Tan thesis aligned with Karpathy wiki pattern + Knowledge Brain System / Nessie export tooling in same batch. Source: X/@garrytan, 2026-05-31

Hermes Desktop

Introducing Hermes Desktop: everything you love about Hermes, now native - @NousResearch | 2026-06-09 | 12,704 likes

Takeaway: Native shell for Hermes Agent (Nous Research). Source: X/@NousResearch, 2026-06-09

Hermes anti-slop operator setup

this is the Hermes setup top 1% operators are using to get rid of AI slop - @EXM7777 | 2026-05-31 | 1,974 likes

Takeaway: Deep-reviewed local artifact now lives on The Eval Loop (Slop Is an Output Problem): the durable lesson is output gating, not more prompt/context tuning. Pair Hermes Agent (Nous Research) with a benchmark, threshold, pass/fail loop, and production failure capture. Source: X/@EXM7777 and local artifact review, 2026-07-03

Hermes Unreal Engine MCP catalog

Unreal Engine's new MCP is now part of the Hermes Agent MCP Catalog. - @Teknium | 2026-06-18 | 102K followers

Takeaway: Promoted to Hermes Agent (Nous Research). Epic's Unreal MCP is an experimental same-machine editor bridge at http://127.0.0.1:8000/mcp, with no auth and serialized game-thread tool calls; Hermes adds catalog install/discovery and MCP server filtering. Treat it as a local creative/editor control pattern, not a remote production-control surface. Source: X/@Teknium and local artifact review, 2026-07-03; Source: Epic Unreal MCP docs, 2026-07-03

NVIDIA SkillSpector for agent skills

NVIDIA built exactly what I needed to secure agent skills — adding as GitHub Action - @dani_avila7 | 2026-05-31 | 498 likes

Takeaway: Supply-chain scanner for skills (928★ repo). Complements Brin - Agent Context Security three-gate chain. Source: X/@dani_avila7, 2026-05-31

Colgate synthetic market research

A toothpaste company has quietly killed the entire market research industry — Colgate published synthetic research at scale - @HowToAI_ | 2026-06-11 | 7,304 likes

Takeaway: Consumer-brand synthetic research as Services-as-Software autopilot — agents replacing panel surveys. Source: X/@HowToAI_, 2026-06-11

Unresolved X Articles (login wall)

@trq212, @RLanceMartin, @nurijanian — linked X long-form articles blocked to anonymous fetch (451). The @mvanhorn row was partially resolved through its self-thread and linked AgentMail-to-Claude-Code source repo, but the long-form article body still needs authenticated re-enrichment if exact article text is required. Source: compiled from enrichment, 2026-06-13; source review, 2026-07-02

---

Timeline

• 2026-05-17 | Initial compilation - 40+ entries from X bookmark AI/agents category. Source: raw/x-bookmarks/bookmarks.jsonl, compiled 2026-05-17
• 2026-06-13 | Newest-55 absorption — polished Jun 2026 sync section; dedicated pages Claude Fable 5, Shopify Quick, Roboflow Supervision, Frontend and Design Skills; updated Hermes Agent (Nous Research), Agent Looping, Security and Review Skills. Source: raw/x-bookmarks/enriched/, User, 2026-06-13
• 2026-06-01 | Absorbed BATCH A (25 enriched bookmarks: AI agents, memory, RAG, security, model announcements). Added entries across memory, skills, infrastructure, multi-agent, design, and new Security/Red-Teaming and Adjacent sections. Created dedicated pages Agent Sandboxes, Learn Harness Engineering, Adversarial Self-Audit Prompts, ZeroEntropy. Source: raw/x-bookmarks/enriched/, compiled 2026-06-01

Batch 151–512 Absorption (74 entries)

Some tips from the skill file that you can feed to your favorite coding agents

Some tips from the https://t.co/RP4FIdUrS8 skill file that you can feed to your favorite coding agents. https://t.co/sHa9i3cRzx - @emilkowalski | Wed 11, Mar | 2,162 likes, 2,776 bookmarks

Takeaway: The bookmark belongs to Animations.dev, Design and Animation, and design-engineering-polish, not a separate AI-agents page. The practical rules are executable UI polish guidance for coding agents: responsive press feedback, natural entry scale, narrow will-change, child-only hover animation, origin-aware popovers, instant subsequent tooltips, 44px tap targets, subtle blur, and touch-safe hover media queries. Source: X/@emilkowalski, 2026-03-11

HOW TO GET THOUSANDS OF VIEWS USING CLAUDE CODE AND REMOTION

HOW TO GET THOUSANDS OF VIEWS USING CLAUDE CODE AND REMOTION one guy created demos that got him thousands of new visitors for his startup here's the exact process: 1\ install remotion (react-based video generation) code becomes a video. thats literally it. 2\ use claude https://t.co/XZOxIjL5qN - @om_patel5 | Sun 15, Mar | 1,132 likes, 2,739 bookmarks

Takeaway: Promoted into Remotion + Claude Code Marketing Demos and refreshed Remotion. Deep video review showed a structured Fyltr product demo (AI form generation, document extraction, smart verification, visualization, filtering, conditional logic), so the durable pattern is Claude Code + Remotion for exact product-UI narrative demos, not generic motion output. Source: X/@om_patel5, 2026-03-15; Source: local video contact sheet, 2026-07-03

Introducing MiniMax M3: The First Open-Weights Model to Combine Three Frontie...

Introducing MiniMax M3: The First Open-Weights Model to Combine Three Frontier Capabilities - Coding & Agentic Frontier: 59.0% SWE-Bench Pro, 66.0% Terminal Bench 2.1, 34.8% SWE-fficiency, 28.8% KernelBench Hard, 74.2% MCP Atlas - MiniMax Sparse Attention scales context to 1M - https://t.co/TF8... - @MiniMax_AI | Mon 01, Jun | 9,978 likes, 2,618 bookmarks

Takeaway: MiniMax M3 is an eval-gated candidate for long-context multimodal agents, not a default routing change. The launch image adds UI/diagram/browser/economic-task benchmarks beyond the text tweet, while the Hugging Face snapshot shows license:other and custom_code, so sandboxed evaluation is mandatory before production harness use. See MiniMax M3. Source: X/@MiniMax_AI, Mon 01, Jun

how I’m building an agent company inside my agency

how I’m building an agent company inside my agency. the structure looks like this: Agency gBrain → Orchestrator Hermes Agent → Department verticals → Specialist agents → Scoped sub-agents gBrain is the company brain. It gets ingested with the data and experience we https://t.co/uzlSrT8FXM - @shannholmberg | Thu 04, Jun | 1,462 likes, 2,532 bookmarks

Takeaway: Promoted to Agent Looping and Knowledge Brain System. The artifact is a real org-architecture diagram: agency brain maintained by human champion + Hermes orchestrator, department verticals, specialist agents, scoped sub-agents, and isolated client pods. The transferable pattern is build one vertical well, fork it to the next client, and keep client memory/approvals isolated. Source: X/@shannholmberg and local visual review, 2026-07-01

HTML-in-canvas is now a first-class primitive in Remotion

HTML-in-canvas is now a first-class primitive in Remotion! It enables new types of effects that were impossible before. https://t.co/fIskOhV47W - @Remotion | Mon 04, May | 2,844 likes, 2,508 bookmarks

Takeaway: Promoted through Remotion and HTML-in-Canvas. The durable agent-workflow signal is that Remotion now has a first-class <HtmlInCanvas> component plus examples/prompts for agent-authored motion effects. The reviewed video shows magnified DOM text, progressive blur over article copy, a John Whitney motion-graphics reference, and a prompt-driven glitch/CRT setup. Current version snapshot: npm remotion latest 4.0.484; docs mark HTML-in-canvas rendering support from v4.0.455. Source: X/@Remotion thread, 2026-05-04; Source: Remotion docs, 2026-06-26; Source: local contact sheet /tmp/xreview-frames/2051235642424631514.jpg, reviewed 2026-07-01

SOMEONE JUST OPEN SOURCED THE MOST COMPLETE CLAUDE CODE SETUP I'VE EVER SEEN

SOMEONE JUST OPEN SOURCED THE MOST COMPLETE CLAUDE CODE SETUP I'VE EVER SEEN 27 agents, 64 skills, and 33 commands. all pre-configured and ready to go planning, code review, fixes, TDD, token optimization. basically everything you'd spend weeks setting up yourself already done https://t.co/fe0Mqhp... - @om_patel5 | Sat 11, Apr | 1,395 likes, 2,454 bookmarks

Takeaway: Promoted into Everything Claude Code - Agent Harness Performance System as earlier adoption evidence for the ECC agent-harness bundle. The local screenshot shows the README surface with language tabs, MIT/security/contributing tabs, 121K-star/16K-fork-era badges, package-download badges, and a visible tree of specialized subagents for planning, architecture, TDD, code review, security review, build-error resolution, docs updates, and language-specific review. Current source authority supersedes those artifact counts: affaan-m/ECC remains at v2.0.0, HEAD 81af4076, with ecc-universal@2.0.0 and ecc-agentshield@1.4.0. Keep ECC as a reference harness, not a wholesale import. Source: X/@om_patel5 and local artifact review, 2026-07-03; Source: GitHub/npm, 2026-07-03

It's a crime that this repo only has under 400 stars

It's a crime that this repo only has under 400 stars. /autoreview and /handoff are literally two of the best agent skills I've ever used. Massive thanks @steipete and @openclaw https://t.co/EWKShtOAhi - @hungv47 | Fri 05, Jun | 1,235 likes, 2,338 bookmarks

Takeaway: Promoted to OpenClaw Agent Skills. Local decision: keep adopted Security and Review Skills and Cloud, Data, and Service Skills, keep Crabbox documented, and defer agent-transcript / session-viewer until privacy and redaction review. The screenshot's 377-star count is historical; current source check shows 854 stars, 51 forks, MIT license, and HEAD 283f069. Source: X/@hungv47, 2026-06-05

🚨BREAKING: Someone just built the open source Loom killer that turns your sc...

🚨BREAKING: Someone just built the open source Loom killer that turns your screen recordings into AI agents. It's called Bloom and it does something Loom never could. Your recordings don't just sit as files. They become queryable, searchable, agent-ready data the moment you https://t.co/EcMOt12hQy - @heyrimsha | Tue 07, Apr | 820 likes, 2,249 bookmarks

Takeaway: Bloom (Agentic Screen Recorder) is the durable route. Local video review confirms the concrete shape: local desktop capture UX plus VideoDB-backed transcription, visual indexing, query, and share. Current source snapshot: canonical repo video-db/bloom, README installer line bloom-2.2.0, public tag check only surfaced v2.1.0, default-branch HEAD 65aa9bf; treat Git tags as incomplete release metadata. Source: X/@heyrimsha, 2026-04-07; Source: GitHub, 2026-07-01

Most important software in my stack:

Most important software in my stack: #1 Tailscale - SSH into all your devices securely - Host websites for use in your home network - Easy high quality VPN #2 VibeProxy - Connect all your AI subscriptions - Single API which will connect to any agent #3 Codex App - Fire - @0xSero | Sat 06, Jun | 2,109 likes, 2,248 bookmarks

Takeaway: Most important software in my stack:

#1 Tailscale

• SSH into all your devices securely
• Host websites for use in your home network
• Easy high quality VPN

#2 VibeProxy

• Connect all your AI su Source: X/@0xSero, Sat 06, Jun

anthropics growth marketer mapped out 4 levels of AI marketing use

anthropics growth marketer mapped out 4 levels of AI marketing use most people sit at level 1, automating what they already do > level 1: automate what you already do (reporting, copy, data pulls) > level 2: use AI as a thinking partner where its better than you > level 3: do https://t.co... - @shannholmberg | Tue 07, Apr | 981 likes, 2,102 bookmarks

Takeaway: Promoted to AI Work Maturity Ladder. Local image review makes the useful ladder explicit: Level 1 automates existing reporting/copy/data pulls, Level 2 uses AI as a thinking partner, Level 3 creates newly economical recurring audits and QA, and Level 4 builds custom systems around owned data/workflows/taste. This is the marketing/operator mirror of No One-Off Work. Source: X/@shannholmberg and local image artifact, 2026-07-03

Ok so @steipete was right

Ok so @steipete was right. Autoreview is by far the best skill I've ever installed. My process used to be branch > PR > CI + @coderabbitai > 100% green > merge Now it's autoreview > merge. - @MatthewGunnin | Thu 04, Jun | 1,081 likes, 2,004 bookmarks

Takeaway: This is adoption proof for the existing Security and Review Skills skill, not a reason to loosen merge gates. The useful signal is the workflow compression: review locally with Autoreview before PR/CI/CodeRabbit-style feedback, then merge only after the normal proof still passes. Keep Autoreview as an advisory closeout gate that finds issues before external review; it does not replace tests, typechecks, source evidence, or reviewer judgment. Source: X/@MatthewGunnin, 2026-06-04

watch openai acquire them in like 6mo

watch openai acquire them in like 6mo https://t.co/rRjap7do4J - @aridutilh | Wed 11, Mar | 2,069 likes, 1,967 bookmarks

Takeaway: Promoted to OpenUI -- Generative UI Language and Runtime. The screenshot shows OpenUI's generative-UI positioning: OpenUI Lang, Playground, API Reference, npx @openuidev/cli@latest create, and a before/after hotel-search example where prose becomes structured cards. Current source checks identify thesysdev/openui as the implementation authority: MIT, HEAD 8c897782, @openuidev/cli@0.1.2, React core packages through @openuidev/react-ui@0.12.1, and Brin caution verdicts for both repo and bundled skill pending deep scan. Use it as a compact streaming generative-UI reference; do not import the skill yet. Source: X/@aridutilh and local artifact review, 2026-07-03; Source: GitHub/npm/Brin, 2026-07-03

Holy shit

Holy shit... someone just open-sourced the cheat code for making AI writing undetectable. It's called stop-slop and it strips every known AI tell from your prose automatically. No rewriting tools. No paraphrasers. No "humanizer" apps. Here's how it works: → A single SKILL.md https://t.co/gS2H6tE... - @heygurisingh | Wed 18, Mar | 867 likes, 1,951 bookmarks

Takeaway: Duplicate Stop Slop signal reviewed against the existing page. The X image 403'd during media repair, but the source still points at the same public skill package; current GitHub state is MIT, HEAD 8da1f03, 13,255 stars, 928 forks, and no public releases/tags. The adoption decision stays unchanged: Security and Review Skills remains Kevin's active de-AI-writing skill; stop-slop is upstream reference material. Source: X/@heygurisingh and GitHub, 2026-07-03

New agent skill: 𝚛𝚎𝚊𝚌𝚝-𝚟𝚒𝚎𝚠-𝚝𝚛𝚊𝚗𝚜𝚒𝚝𝚒𝚘𝚗𝚜

New agent skill: 𝚛𝚎𝚊𝚌𝚝-𝚟𝚒𝚎𝚠-𝚝𝚛𝚊𝚗𝚜𝚒𝚝𝚒𝚘𝚗𝚜 Add React <𝚅𝚒𝚎𝚠𝚃𝚛𝚊𝚗𝚜𝚒𝚝𝚒𝚘𝚗> animations to any React app. Also covers how Next.js can natively integrate them. • Animate elements across navigations • Slide pages forward and back • Smooth loading https://t.co/2uTynQL3lt - @nextjs | Thu 02, Apr | 2,177 likes, 1,874 bookmarks

Takeaway: Thread: New guide: implement each animation from the video with React <𝚅𝚒𝚎𝚠𝚃𝚛𝚊𝚗𝚜𝚒𝚝𝚒𝚘𝚗> in Next.js. Demo video attached. Source: X/@nextjs, Thu 02, Apr

Claude Cowork is a literal CHEAT CODE for SEO

Claude Cowork is a literal CHEAT CODE for SEO. I could take ANY local service business to $100k/month in 90 days using just Claude Cowork. Here’s exactly how I'd do it: https://t.co/ub9GkjG2L2 - @bloggersarvesh | Sat 14, Mar | 786 likes, 1,830 bookmarks

Takeaway: Promoted into Writing and Content Skills plus ai-seo / programmatic-seo: use browser agents for local-service keyword research, competitor page extraction, Google Business Profile completeness, service-page locality signals, review-response drafts, schema checks, and rank/citation monitoring. The attached images are Claude and Google Business Profile logos; the durable value is the workflow checklist. Source: X/@bloggersarvesh, Sat 14, Mar

Nessie just became the best way to get all your existing context, memory and ...

Nessie just became the best way to get all your existing context, memory and history from ChatGPT, Perplexity, and Gemini into all the other places you have memory, and also get it into OpenClaw/Hermes Agent. Their OpenClaw and MCP servers are ace. https://t.co/0L1JaBlTvg - @garrytan | Thu 11, Jun | 1,140 likes, 1,795 bookmarks

Takeaway: Source-reviewed into Nessie. The OpenClaw plugin writes Nessie's hosted MCP server into OpenClaw config after email-code or API-key setup, so this is a concrete memory-portability bridge between AI transcript history and local agent harnesses. Current source snapshot: nessielabs/nessie-openclaw tag v0.1.13; hosted MCP server https://mcp.nessielabs.com/mcp. Source: X/@garrytan, 2026-06-11; Source: GitHub nessielabs/nessie-openclaw, 2026-07-04

T3 Code is now available for everyone to use

T3 Code is now available for everyone to use. Fully open source. Built on top of the Codex CLI, so you can bring your existing Codex subscription. https://t.co/XUXUo7cfPn - @theo | Sat 07, Mar | 4,732 likes, 1,753 bookmarks

Takeaway: Deep review promoted the thread into T3 Code. T3 Code is a Codex CLI-based, open-source, parallel-agent operator surface: useful as evidence that the harness value sits in orchestration, scannability, and task state, not only in the model call. Current authority: t3@0.0.28 latest, same-day 0.0.29-nightly.20260701.697, and pingdotgg/t3code HEAD 7b9eef7ac29f9d4819c6411dfb1c5f04fef50264. Source: X/@theo, Sat 07, Mar; Source: npm/GitHub, 2026-07-01

Introducing the gradient border plugin

Introducing the gradient border plugin Familiar syntax, CSS-only, works everywhere. Easiest way for agents (and humans) to add border gradients via tailwind. https://t.co/xqy9mZBKpO - @flornkm | Wed 29, Apr | 1,366 likes, 1,719 bookmarks

Takeaway: Gradient Border Plugin (Tailwind) is a Tailwind v4 CSS-only gradient-border utility. Deep review preserved the demo video/image and refreshed the package snapshot: gradient-border-plugin@1.1.3, repo HEAD 51df2f9, 148 stars. The useful signal is agent-friendly syntax for a visual effect that otherwise becomes bespoke mask-composite CSS. Source: X/@flornkm, Wed 29, Apr; Source: npm registry and GitHub API, 2026-07-02

Our bootstrapped SaaS just outranked our 6 VC backed competitors in under 48...

Our bootstrapped SaaS just outranked our 6 VC backed competitors in under 48 hours. ChatGPT and AIOs are actively sh*t talking them and recommending us Here's exactly how we did it: https://t.co/5TXR9GqMSO - @RobHoffman_ | Fri 13, Mar | 566 likes, 1,631 bookmarks

Takeaway: Promoted into Writing and Content Skills plus ai-seo / programmatic-seo: exact-match comparison/review pages can win Google AI Overview and AI-search recommendations when they combine firsthand product testing, external review language, dated pricing, honest tradeoffs, extraction-friendly verdicts, and follow-up checks across Google AIO/AI Mode, ChatGPT, Perplexity, and Gemini. The reviewed image shows a Kleo vs Stanley Google result with Kleo cited in the AI Overview and organic results. Source: X/@RobHoffman_, Fri 13, Mar

just updated ui-skills:

just updated ui-skills: It's now a collection of the best skills for design engineering https://t.co/kRYdArErER - @Ibelick | Tue 28, Apr | 1,017 likes, 1,537 bookmarks

Takeaway: Promoted through UI Skills Pack. The video shows UI Skills as a broad design-engineering registry, not a small personal pack: cards for baseline UI, shadcn, transitions, React Doctor, interface/interactions, OKLCH, animation principles, and make-interfaces-feel-better. Current source check: 132 registry entries, ui-skills@0.2.3, GitHub HEAD 46ef54de56db504c7a568ccddb6acd6ef8dc8815. Keep it as an external discovery shelf, not an automatic install list. Source: X/@Ibelick and local video review, 2026-07-03

Claude + GSC SEO Hack:

Claude + GSC SEO Hack: 1. Go to Google Search Console 2. Pick 1 page thats doing well in the last 90 days 3. Filter and download all the queries data for that page 4. Upload this CSV to Claude with the following prompt: "I wrote this article, its getting some traffic - [LINK] https://t.co/ZGOXz44R... - @ayushtweetshere | Sat 14, Mar | 690 likes, 1,493 bookmarks

Takeaway: Thread: check out my newsletter for more tactical content - Source: X/@ayushtweetshere, Sat 14, Mar

Deep-review update: Promoted to Writing and Content Skills as a Google Search Console query-refresh loop: pick a page with traction, export 90-day page-filtered query data, group queries into satisfied and missing intent, then refresh headings, answer blocks, FAQ/schema, and internal links before measuring the next crawl window. The local screenshot is useful only as proof of the GSC input shape; the durable value is the CSV-to-refresh workflow. Source: X/@ayushtweetshere, 2026-03-14

Introducing the Hermes Agent Profile Builder

Introducing the Hermes Agent Profile Builder You can now build a complete profile in the dashboard with full control over identity/name/description, model/provider, built-in + optional skills, skills-hub installs, and MCP servers in one easy flow https://t.co/bHvqczZYmz - @NousResearch | Wed 10, Jun | 2,983 likes, 1,432 bookmarks

Takeaway: Updated Hermes Agent (Nous Research) with the Profile Builder as a concrete capsule boundary. The demo flow is Identity → Model → Skills → MCPs → Review, which is the exact structure Kevin's agent brain should preserve: public profile/config/skills/connectors are packageable, while private memory, sessions, secrets, and local files stay outside the profile. It also validates a UI pattern for new-agent boot: users should compose identity, model policy, skill routes, and MCP access in one reviewable flow before activation. Source: X/@NousResearch, 2026-06-10; Source: Hermes web dashboard docs, 2026-06-30

Hermes learns TouchDesigner

I didn't touch TouchDesigner myself. Hermes agent learned it from scratch and built this: → navigated my desktop with computer use → figured out how to connect to TouchDesigner → read my reference images → iterated on the art with me in a self-learning loop → then saved - @whosamberella | 2026-06-13 | 561 likes, 729 bookmarks

Takeaway: Promoted to Hermes Agent (Nous Research) and Computer Use and Browser Automation Patterns. The durable signal is a creative GUI learning loop: Hermes navigates the desktop, discovers the TouchDesigner connection path, reads reference images, iterates with a human, and saves output. The reviewed contact sheet also shows Hermes desktop/provider settings and Minimax M3/VLM setup, so treat this as evidence for agent-operated creative tools, not a new local skill yet. Source: X/@whosamberella, 2026-06-13; Source: local video contact sheet, 2026-07-02

bookmark this immediately

bookmark this immediately. someone just open sourced the methodology that changes how ai agents write code. 40,000 stars on github in days. it's called superpowers and if you're using claude code or codex without this, your so cooked. don't fall behind. - @iruletheworldmo | Tue 17, Mar | 710 likes, 1,426 bookmarks

Takeaway: Duplicate adoption evidence for Superpowers. Deep review refreshed the current source snapshot to v6.1.1 / 246K+ GitHub stars and kept the local decision unchanged: use Kevin's existing skills-first rules and Superpowers-style workflow rather than creating another executable skill. Source: X/@iruletheworldmo, 2026-03-17; Source: GitHub obra/superpowers, 2026-07-04

IT'S SO OVER

IT'S SO OVER... Motion designers are cooked. This AI agent made this video in 17 mins, only did 2 changes. https://t.co/4UWCD9HRDk - @chddaniel | Tue 09, Jun | 983 likes, 1,226 bookmarks

Takeaway: Promoted to Shipper.now. The video is source-critical: generated product-motion frames sell a chat-to-app/game builder, with "You don't [need] Design," "You don't have $2 billion dollars," a Games-tab prompt, a generated block-survival game preview, and behind-the-scenes build framing. Shipper has no public package version; hosted-service snapshot checked 2026-07-03. Source: X/@chddaniel, 2026-06-09; Source: shipper.now, 2026-07-03

With the new scrambleText() feature, you can control reveal direction, create...

With the new scrambleText() feature, you can control reveal direction, create cursor effects, apply perturbation and easing, and chain it in keyframes and timelines like any other Anime.js animation property. It's crazy how just a few lines of code can create such visually https://t.co/pi8BriIiu8 - @JulianGarnier | Wed 29, Apr | 1,465 likes, 1,225 bookmarks

Takeaway: Deep review routes this to Text Animations Collection (animate-text), X Bookmarks: Design & UI (Feb–Jun 2026), and the animejs skill. Use Anime.js scrambleText() when the product needs terminal-style text reveal, randomized copy transitions, cursor-led decoding, or direction-aware character reveal inside an Anime.js project. Source: X/@JulianGarnier, 2026-04-29; Source: Anime.js docs, 2026-07-02

Just want to make this clear:

Just want to make this clear: We didn't make Hermes Agent to be a "starts with nothing, you work it all out" agent. This is not the minimalist, start from nothing, agent. We want Hermes to work out of the box for most people. So you aren't spending weeks just getting the https://t.co/3biQNjGeO... - @Teknium | Sun 31, May | 4,200 likes, 1,211 bookmarks

Takeaway: Deep review promoted Hermes Profile Distributions into Hermes Agent (Nous Research) and Agent Harness Capsules. The useful pattern is a versioned public agent package: SOUL/config/skills/cron/MCP/files in git, with memories, sessions, API keys, and secrets outside the distribution boundary. Source: X/@Teknium, 2026-05-31; Source: Hermes docs

More suggestions:

More suggestions: Vercel, Anduril, Mercury, Stripe, ElevenLabs, Profound, Anthropic, Mintlify, Databricks, SpaceX, Elise AI, Browserbase, Palantir, Notion, Adaptive Security, Slash, Standard Metrics, Superpower, Nozomio, Foam, Lovable, Polymarket, Starcloud, Linear, Boltnew - @benln | Fri 10, Apr | 1,204 likes, 1,167 bookmarks

Takeaway: Routed to Ben Lang and Career Profile. This is a future-founder company-targeting list, not an AI-agent tool list: Vercel, Anduril, Mercury, Stripe, ElevenLabs, Profound, Anthropic, Mintlify, Databricks, SpaceX, Elise AI, Browserbase, Palantir, Notion, Adaptive Security, Slash, Standard Metrics, Superpower, Nozomio, Foam, Lovable, Polymarket, Starcloud, Linear, and Bolt.new. Source: X/@benln, 2026-04-10

Obsidian

Rewrites Obsidian vault pages using AI agents https://t.co/tWem59TciH https://t.co/AHzCO6mZfQ - @tom_doerr | Sun 31, May | 680 likes, 1,023 bookmarks

Takeaway: See Obsidian Second Brain; deep review refreshed the current six-platform, 44-command release snapshot and preserved the older v0.8 image drift. Source: X/@tom_doerr, Sun 31, May

Matt has been shipping some incredible products lately

Matt has been shipping some incredible products lately... you should definitely equip your agents with: - /last30days - PrintingPress - Agent Cookie completely different experience with Hermes especially - @EXM7777 | Mon 01, Jun | 612 likes, 992 bookmarks

Takeaway: Promoted to Matt Van Horn, Security and Review Skills, Printing Press - Agent-Native CLIs from a Single Prompt, and Agent Cookie - Session State Sync for the Agent's Second Mac as a stack-composition signal for Hermes: research layer, generated service CLIs, and auth/session sync together change the agent operating experience. Source review recorded upstream drift for /last30days and current source snapshots for the three tools. Source: X/@EXM7777, 2026-06-01

a love letter to the thousands of people who downloaded this over last 48 hrs,

a love letter to the thousands of people who downloaded this over last 48 hrs, run this: › npm install -g fieldtheory@latest › ft sync --gaps (grabs threads, RTs) bonus: › ft wiki (creates md wiki, 🕗) › ft md (exports md) › ft skill install (/fieldtheory) - @andrewfarah | Wed 08, Apr | 619 likes, 966 bookmarks

Takeaway: Field Theory CLI (Fieldtheory) is the self-custody X bookmark collector behind this wiki's bookmark pipeline. Deep review refreshed the command surface (ft sync --gaps, ft wiki, ft md, ft skill install) and current source snapshot: fieldtheory@1.3.22, MIT repo HEAD 4b4a0f6, 1,938 stars, 201 forks, no release tags returned. Source: X/@andrewfarah, 2026-04-08; Source: npm/GitHub, 2026-07-04

Nano Banana 2 is beyond cracked

Nano Banana 2 is beyond cracked. There is not a static ad out there that can't be one-shot with Claude + Higgsfield + NB2 https://t.co/JljwKhGnfX - @alexgoughcooper | Wed 11, Mar | 567 likes, 839 bookmarks

Takeaway: Nano Banana 2 is beyond cracked.

Deep review routes this to Higgsfield MCP as static-ad composition evidence: four local image artifacts show DTC-style ad comps with product packshots, claims, CTA, social proof, and brand-specific art direction. Do not treat this bookmark as primary source authority for Nano Banana 2 model specs. Source: X/@alexgoughcooper and local image review, 2026-07-03

Garry is kinda correct here, but is oversimplifying memory

Garry is kinda correct here, but is oversimplifying memory. Harrison (the author of the original article) makes a very good point but also makes memory sound easier than it is. (before reading this article, note that I wrote down my thoughts and then passed it through Claude - @michael_chomsky | Sun 12, Apr | 395 likes, 754 bookmarks

Takeaway: Promoted to Agent Memory Patterns as the counterweight to "memory is markdown." The source-reviewed thread says memory remains a tradeoff system: collaboration, observability, update policy, benchmarks, and database-like failure modes still matter even when portable files own the durable write path. See Claude-Mem for one implementation reference. Source: X/@michael_chomsky and self-thread, 2026-07-04

If you’re just looking to create Ecom pages from Claude to Shopify here is an...

If you’re just looking to create Ecom pages from Claude to Shopify here is an easier method Without needing to setup Shopify MCP and custom app -Create a repo folder for your brand info and skills -Install Shopify CLI in terminal (the dev tool use to create and modify theme - @oliviercroguy | Thu 12, Mar | 313 likes, 701 bookmarks

Takeaway: Promoted into Service CLI Coverage as a first-party CLI shortcut: for simple agent-authored Shopify pages, start with brand/theme context files plus Shopify CLI before inventing a custom Shopify MCP app. No versionable tool claim was present beyond the existing shopify CLI route. Source: X/@oliviercroguy, 2026-03-12

Claude Mythos / Fable 5 solved motion graphics with HTML

Claude Mythos / Fable 5 solved motion graphics with HTML! this @HyperFrames_ video was completely one-shotted by Mythos btw https://t.co/onFKDGwx0m - @Miguel07Code | Tue 09, Jun | 960 likes, 685 bookmarks

Takeaway: Deep review promoted this into HyperFrames as agent-run reproducibility evidence: the thread linked the Claude connector, repo, source session, and launch-code example, and the current source snapshot is hyperframes@0.7.21 / GitHub release v0.7.21 as of 2026-06-30. Source: X/@Miguel07Code, 2026-06-09; Source: GitHub/npm, 2026-06-30

YC is partnering with over a dozen companies to offer students attending YC e...

YC is partnering with over a dozen companies to offer students attending YC events free credits for AI developer tools! The YC AI Student Starter Pack includes over $20k in Azure & AWS credits, $5k+ in ChatGPT, Claude, and Grok credits, and more! https://t.co/0QEerEDoqf https://t.co/HsyMfbShqc - @ycombinator | Fri 16, Jan | 997 likes, 683 bookmarks

Takeaway: Promoted to YC AI Stack. This is a student-founder credit bundle for AI developer tools, useful as prototype budget after eligible YC university events but not an architecture dependency. Source: X/@ycombinator, 2026-01-16; Source: Y Combinator, 2026-01-16

get your product mentioned in chatgpt or ai search

get your product mentioned in chatgpt or ai search! https://t.co/xRKiuvw2qS - @jia_seed | Tue 03, Mar | 838 likes, 647 bookmarks

Takeaway: Promoted to Jam (SpreadJam). Local video review identifies the product as Jam/SpreadJam: a growth-channel agent tool that runs an SEO/GEO audit, reads the selected GitHub repo/site, finds quick wins and content gaps, and presents PR-style changes. Source review added the current SpreadJam site/docs, wespreadjam/jam-nodes, and @jam-nodes/playground@0.1.2. Source: X/@jia_seed, 2026-03-03; Source: SpreadJam site/docs/GitHub/npm, 2026-07-02

ガチで助かる。GitHub 15,000+ Stars。

ガチで助かる。GitHub 15,000+ Stars。 APIキー不要でAIエージェントがネット全体を読めるようになりました。名前は Agent Reach。 ・Twitter/X、Reddit、YouTube、GitHub など15+プラットフォームに対応 ・API費用ゼロ（CLIベースのオープンソース） ・ほぼコマンド1つで導入できる https://t.co/A30A46g26U - @so_ainsight | Tue 07, Apr | 389 likes, 636 bookmarks

Takeaway: Security and Review Skills is an installer/doctor/channel-router capability layer, not a runtime wrapper: it registers a skill, wires upstream CLIs/channels for web, X, YouTube, GitHub, Reddit, Bilibili, XiaoHongShu, LinkedIn, RSS, and Exa, then lets agents call those channels directly. Current source snapshot is v1.5.0, MIT, HEAD e87bcdb17713c24513f6151d508f701b87dfd797, 48,254 stars, and 3,836 forks; keep the cookie-based routes behind dedicated accounts, safe mode, dry run, and config review. Source: X/@so_ainsight thread, 2026-04-07; Source: Agent Reach README

introducing

introducing https://t.co/Qnkj0R2T15 replace "hub" with "inspect" on any github url and chat with any repo. - pi agent that runs in your browser. - just-bash with a VFS built on top of the github API - everything stays local. turns out pi + just-bash is all you need. cc https://t.co/UWI4Slv7A8 - @dinnaiii | Thu 09, Apr | 529 likes, 612 bookmarks

Takeaway: Promoted through Gitinspect. The reviewed video shows the core interaction: replace github.com/chenglou/pretext with gitinspect.com/chenglou/pretext, then ask repo-structure/testing questions through an in-browser chat backed by a read-only GitHub VFS. Current source check: AGPL-3.0, 61 stars, 11 forks, HEAD 0f30313, no public tags. Source: X/@dinnaiii, 2026-04-09; Source: local contact sheet /tmp/xreview-frames/2042257371645247716.jpg, reviewed 2026-07-01

the cost of shipping code went to zero

the cost of shipping code went to zero taste didn't but "taste" sounds mystical and unfixable, so nobody teaches it. here's the unmystical version: taste is just an eval you haven't written down yet how you choose what to measure is what matters 1/8 - @sdianahu | Thu 04, Jun | 624 likes, 566 bookmarks

Takeaway: Existing The Eval Loop (Slop Is an Output Problem) owns this thread. Diana Hu's durable point is that product taste becomes operational when teams turn real failure traces into a domain-specific eval taxonomy: generic benchmarks ask whether a model is broadly smart, while product evals ask whether it did the customer job. The thread's useful method is open coding plus axial coding over 1,000+ traces, converting failure modes into customer-value metrics; no new page needed. Source: X/@sdianahu thread, 2026-06-04

💻 «UI Skills» by @ibelick

💻 «UI Skills» by @ibelick A set of skills to polish interfaces built by agents. → https://t.co/KTGBB1mMqY More on → https://t.co/7fk3BqVVP6 🏴 https://t.co/eR12dfc6dJ - @remvze | Tue 24, Mar | 447 likes, 538 bookmarks

Takeaway: UI Skills Pack is now the active external UI-skill discovery shelf; the live registry has 112 entries and routes through Kevin's local design skills first. Source: X/@remvze, Tue 24, Mar

distilling my design aesthetic into a Claude skill — so we can craft design s...

distilling my design aesthetic into a Claude skill — so we can craft design systems together that fits the vibe. since I entered the field, I've spent years handcrafting every interface myself. loved every second of it. but vibe coding unlocked a whole new world for human-ai https://t.co/dS6QczTCUh - @fleeting_land | Tue 12, May | 349 likes, 523 bookmarks

Takeaway: Promoted to Taste Skill Framework (Leonxlnx/taste-skill) as a reference-only personal aesthetic skill pattern. The reviewed demo shows a skill producing named aesthetic modes, principle pages, a motion/gallery surface, and an image-led "Aesthetic Mirror"; useful for taste encoding, but not imported as an active skill without source or an install route. Source: X/@fleeting_land and local video review, 2026-07-03

The hardest thing about agents and backends is durability

The hardest thing about agents and backends is durability. @workflowsdk fixes this. That LLM you're calling will go down. That service will rate limit you. That database will unexpectedly slow down. You will get paged 💀 I've been looking for a unicorn for a decade. I https://t.co/vkvgBMpk... - @rauchg | Thu 16, Apr | 615 likes, 478 bookmarks

Takeaway: Promoted to Vercel Workflows. The local image is a workflow trace, not just a launch card: generateBirthdayCard completed after 1m48s with expiry/storage metadata, trace/events/streams tabs, a waterfall span, a hook wait, and a sleep segment. Current source snapshot: Vercel docs install workflow; npm has workflow@4.5.0 latest and 5.0.0-beta.26 beta. Source: X/@rauchg, 2026-04-16; Source: Vercel docs/npm registry, 2026-07-03

I feel like this could be turned into a cool hero effect for a component libr...

I feel like this could be turned into a cool hero effect for a component library site. Where you move your cursor to tear between the design and the code.... 🤔 https://t.co/dEZ95guoka - @mattrothenberg | Thu 09, Apr | 811 likes, 455 bookmarks

Takeaway: Rerouted to Creative Web Interaction References as a cursor-controlled design/code tear hero: a vertical blue/orange seam reveals code on one side and the rendered issue form on the other, useful for component-library source/result storytelling. Source: X/@mattrothenberg and local video contact sheet, 2026-07-03

We just shipped 𝙷𝚊𝚛𝚗𝚎𝚜𝚜𝙰𝚐𝚎𝚗𝚝, a unified abstraction to orchestrat...

We just shipped 𝙷𝚊𝚛𝚗𝚎𝚜𝚜𝙰𝚐𝚎𝚗𝚝, a unified abstraction to orchestrate and integrate any agent’s “brain” into your app. @aisdk now frees you from both model and agent lock-in. (And it doesn’t just get you portability, it’s also delightful to use ofc!) - @rauchg | Fri 12, Jun | 508 likes, 431 bookmarks

Takeaway: Promoted to AI SDK HarnessAgent. Guillermo's row is the product thesis version of the Vercel Developers API row: app teams should be able to swap agent harnesses like they swap model providers, while keeping AI SDK-compatible streams and sandboxed sessions. Current package snapshot already lives on the canonical page. Source: X/@rauchg, 2026-06-12; Source: Vercel changelog, 2026-06-12

Anthropic Claude Code skill portfolio taxonomy

ANTHROPIC JUST DROPPED A MASSIVE GUIDE on how they use claude code skills internally tl;dr for people who don't want to read the whole thing: 9 types of skills worth building: > 1. library & API reference - teach claude your internal tools and SDKs > 2. product verification - - @om_patel5 | Wed 18, Mar | 212 likes, 392 bookmarks

Takeaway: Source-reviewed into Agent Skill Bundles, Claude Code Harness, and Agent Harness Capsules. The useful signal is a nine-part portfolio audit for skill libraries: reference, verification, data, team process, scaffolding, review, CI/CD, runbooks, and infrastructure ops. Official Anthropic docs are the runtime authority: skills are on-demand folders of instructions, scripts, templates, and resources, not just longer prompts. Source: X/@om_patel5, 2026-03-18; Source: Anthropic docs, 2026-07-04

The skills​

The skills​.sh API is now generally available. Power your agents, applications and platforms with access to over 600,000 skills ↓ https://t.co/wOXUhbFmxX - @vercel_dev | Fri 05, Jun | 404 likes, 377 bookmarks

Takeaway: Source-reviewed into Skills.sh - Agent Skills Registry. The API lets Vercel-authenticated products query 600,000+ skills, fetch details/security audits, and avoid long-lived registry secrets by using project OIDC tokens. Current CLI/package snapshot checked during review: skills@1.5.14. Source: X/@vercel_dev, 2026-06-05; Source: Vercel changelog and npm registry, checked 2026-07-04

this skill might sharpen ur designs some more

this skill might sharpen ur designs some more https://t.co/C1ojUaCSFo https://t.co/lSPKGPk6hY - @dhasandev | Sun 15, Mar | 208 likes, 347 bookmarks

Takeaway: Candidate only; routed to Design Methodology Atlas, not imported as an active skill. product-design-critic has a strong workflow for product judgment - goal -> JTBD -> primary surface -> supporting context -> critical states -> trust/governance -> recommendation with tradeoffs - but local coverage already exists in Frontend and Design Skills, Frontend and Design Skills, design-engineering-polish, and taste routes. Gist snapshot checked 2026-07-02: one file, no companion references present, raw file e145e518e59ce9222f54075f0407e9a5ff182a1d, version 62b3ef3e8bd6f2300702c707ab5a9722fe871f9e. Source: X/@dhasandev, 2026-03-15; Source: GitHub Gist API, 2026-07-02

i built vault, a self-improving knowledge base that can index your entire lif...

i built vault, a self-improving knowledge base that can index your entire life (personal mini agi) it updates while you sleep, supports 50+ integrations like notion, snowflake, datadog, and more when you connect a source (like stripe), all the data becomes a filesystem built https://t.co/6wIO1GyQ... - @arlanr | Mon 13, Apr | 310 likes, 313 bookmarks

Takeaway: Promoted to Nia Vault. The self-thread links resolve to trynia.ai and the Vault docs; source review confirmed npx nia-wizard@latest maps to nia-wizard@0.1.34. The reviewed video shows the vault UI, terminal indexing, generated markdown pages, Vault Schema, workflow logs, and docs/CLI loops. Source: X/@arlanr, 2026-04-13; Source: local video artifact review and npm registry, 2026-07-02

People obsessed with engineering, computer science, robotics, AI, AGI, machin...

People obsessed with engineering, computer science, robotics, AI, AGI, machine learning, deep learning, neural networks, LLMs, SLMs, transformers, RAG systems, vector databases, agents, automation, DevOps, MLOps, distributed systems, backend architecture, scalable systems, Linux, - @divaagurlxw | Sun 31, May | 736 likes, 312 bookmarks

Takeaway: Reviewed as a parked AI/engineering resource teaser. The root text is a broad interest list, the self-thread only adds "Free will Posting," and the artifact reply resolves to another X status that Jina could not fetch during review. No durable tool, skill, or concept promotion until the linked source becomes accessible or repeats elsewhere. Source: X/@divaagurlxw, 2026-05-31; Source: enriched source review, 2026-07-03

i’m using the superpowers skill for almost everything i’m working on

i’m using the superpowers skill for almost everything i’m working on it’s great - @gavinpurcell | Sun 15, Mar | 181 likes, 288 bookmarks

Takeaway: i’m using the superpowers skill for almost everything i’m working on

it’s great Source: X/@gavinpurcell, Sun 15, Mar

This feels like the future arrived early

This feels like the future arrived early. Your Meta Ray-Ban glasses just became a voice-controlled AI agent that sees what you see, hears what you say, and completes tasks in the real world without you ever touching your phone. It's called VisionClaw and someone open-sourced https://t.co/GLBwPsUdz... - @ihtesham2005 | Sat 28, Mar | 151 likes, 230 bookmarks

Takeaway: Promoted to VisionClaw and cross-linked to OpenClaw Harness, Computer Use and Browser Automation Patterns, and Agent Product Surface. Primary-source review confirms the architecture is Meta Wearables DAT SDK + Gemini Live + optional OpenClaw gateway, with phone-camera fallback and WebRTC POV streaming; GitHub reports license NOASSERTION, so treat it as a prototype under Meta Wearables terms rather than a generic open-source dependency. Source: X/@ihtesham2005, 2026-03-28; Source: Intent-Lab/VisionClaw, 2026-07-04

Figma → Create the inspiration visuals

Figma → Create the inspiration visuals Claude Design → Initial build Claude Code → Iterate and polish https://t.co/RUz3FtsgSR - @Oluwaphilemon1 | Wed 13, May | 222 likes, 218 bookmarks

Takeaway: See AI-Native Design Patterns. Durable workflow: Figma creates the inspiration visuals, Claude Design does the initial build, and Claude Code iterates/polishes the implementation. The reviewed video shows a dark motion-heavy portfolio concept, so the insight is visual-first agent constraint rather than a new standalone tool. Source: X/@Oluwaphilemon1, 2026-05-13

Best companies for future founders to work at these days: Ramp, Cursor, OpenAI

Best companies for future founders to work at these days: Ramp, Cursor, OpenAI. Where else? - @benln | Thu 09, Apr | 537 likes, 200 bookmarks

Takeaway: Reviewed and routed out of the AI-tools hub to Ben Lang and X Bookmarks: Career & Business (Mar–May 2026). This is company-targeting evidence for future-founder learning density, not an agent tool: Ramp, Cursor, and OpenAI are the starting set, with the follow-up list preserved on Ben's page. Source: X/@benln, reviewed 2026-07-04

‼️ YOU ASKED FOR THE SEO’AUCE

‼️ YOU ASKED FOR THE SEO’AUCE Here's my COMPLETE PLAYBOOK Step 0. Your OpenClaw Agent pulls real data from Google Search Console (enable Google API) From there: 1. Sitemaps generated and submitted automatically 2. Unique blog posts and landing pages written and published - - @woocassh | Tue 10, Mar | 103 likes, 176 bookmarks

Takeaway: Hub-level SEO automation pattern. The bookmark's useful shape is "agent pulls real Search Console data, then proposes/ships sitemap, blog, landing-page, and indexing work." Active execution still routes through Writing and Content Skills, Security and Review Skills, Security and Review Skills, Jam (SpreadJam), or NotFair (SEO and Ads Claude Skills) depending on whether Kevin needs methodology, technical audit, productized growth workflow, or account-connected marketing ops. No new skill: the row has no repo/docs/source artifact beyond the tweet. Source: X/@woocassh, 2026-03-10

Design with gpt-images 2

Design with gpt-images 2.0 web skill https://t.co/iS4aZYb0aC - @LexnLin | Sat 02, May | 282 likes, 172 bookmarks

Takeaway: Taste Skill Framework (Leonxlnx/taste-skill) source-review row. The self-thread resolves to imagegen-frontend-web/SKILL.md; the local video shows the "Forma" site split into horizontal section references. Treat as a reference-only skill candidate after Brin caution, not an adopted executable skill. Source: X/@LexnLin, 2026-05-02

@karpathy thanks Andrej - good feedback & more improvements coming

@karpathy thanks Andrej - good feedback & more improvements coming! writes are more sensitive to spam/abuse especially as agents are now the new norm so we currently set the price higher but reviewing for balance. we’re also actively improving docs and revamping to be far more agent - @chrisparkX | Sun 05, Apr | 384 likes, 127 bookmarks

Takeaway: Promoted into X-Agent. The source thread marks X API's agent-facing surface: Pay-Per-Use, XMCP/xurl, official Python and TypeScript XDKs, API Playground, Docs MCP at https://docs.x.com/mcp, and X MCP at https://api.x.com/mcp through the xurl mcp bridge. Current snapshot: @xdevplatform/xurl@1.2.2, @xdevplatform/xdk@0.5.0, xdevplatform/xmcp HEAD 63d3436. Source: X/@chrisparkX, 2026-04-05; Source: X docs/npm/GitHub, 2026-07-04

why are agents not making money rn

why are agents not making money rn - @itsCathyDi | Fri 06, Mar | 250 likes, 37 bookmarks

Takeaway: Promoted to Agent Unit Economics. Agent monetization needs a paid workflow outcome and margin after model calls, sandbox time, retries, support, and human review; model capability alone is not a business model. Source: X/@itsCathyDi, 2026-03-06

you should cancel your amex gold card

you should cancel your amex gold card. don't believe me? try it out for yourself to see what cards you're overpaying for with this new skill powered by https://t.co/AT3ywxH4tQ, claude code, + @browserbase https://t.co/EkARyjlm5z - @KenzieMac_Dev | Fri 05, Jun | 36 likes, 26 bookmarks

Takeaway: Deep review promoted the browse.sh/NerdWallet demo into Browserbase Browse.sh (Open Web Catalog) and Browse.sh (Browserbase Skills). The skill is useful as a read-only, site-specific structured extraction pattern, but remains a candidate rather than an active Kevin skill because it is third-party, domain-specific, and finance-adjacent. Source: X/@KenzieMac_Dev, 2026-06-05; Source: browse.sh skill page

new recipe to cook w/

new recipe to cook w/ - a lil bun - touch of ts7/tsgo - opinionated oxlint/oxfmt cfg - focused unit tests - agentic hooks depending on project, you can validate your entire repo in ~250-500ms. id like to experiment with running these checks on every loc changed - @rustydotwtf | Tue 10, Mar | 12 likes, 25 bookmarks

Takeaway: Promoted into Ultracite & Fast Validation Recipe as a text-only source for the sub-500ms validation recipe: Bun, TS7/tsgo, oxlint/oxfmt, focused unit tests, and agentic hooks. No new tool or skill was created because the durable route is the existing Ultracite/fast-validation stack; the useful decision is to run checks on every changed line or small edit batch when the repo is fast enough. Source: X/@rustydotwtf, 2026-03-10

had a lot of fun chatting about:

had a lot of fun chatting about: > my fundraising philosophy after raising 11M > common mistakes founders in SF make > why agent security isn’t solved yet (what openclaw is doing wrong) > the best book i read in 2026 full convo: https://t.co/UiV5vkkbz9 - @itsCathyDi | Tue 24, Mar | 30 likes, 17 bookmarks

Takeaway: Source-reviewed into Cathy Di. Treat the episode row as bounded public-positioning evidence: Cathy foregrounded fundraising philosophy, SF founder mistakes, and the unsolved agent-security/deployment layer; the Spotify shell was available, but no full transcript was recovered. Source: X/@itsCathyDi and Spotify episode page, 2026-03-24

99% of useful agents haven’t been built yet

99% of useful agents haven’t been built yet. i call this the ‘imagination chasm’ https://t.co/U8YX9fwCmW - @itsCathyDi | Tue 17, Mar | 33 likes, 15 bookmarks

Takeaway: Promoted to Imagination Chasm. The self-thread resolves to the Open AGI YouTube recording "Arena Opening Day: Cohort 0"; the local X thumbnail is visually blank, so keep this as source/thread evidence rather than image evidence. Durable thesis: most agent value is still blocked by people not being able to picture concrete agent workflows. Source: X/@itsCathyDi, 2026-03-17; Source: YouTube oEmbed, checked 2026-07-03

Everyone claims they’re building agents

Everyone claims they’re building agents. Nobody agrees on what an agent actually is. We think the definition should be simple: Agents := Model(s) + Tool(s) More on how we're thinking about this ↓ https://t.co/kqsjIONJY4 - @dedaluslabs | Tue 24, Mar | 14 likes, 8 bookmarks

Takeaway: Everyone claims they’re building agents. Nobody agrees on what an agent actually is.

We think the definition should be simple:

Agents := Model(s) + Tool(s)

More on how we're thinking about this ↓ Source: X/@dedaluslabs, Tue 24, Mar

how do we solve the imagination chasm for agents

how do we solve the imagination chasm for agents? > taking legacy paradigms and finding where automation can be applied in novel ways. to visualize this, we flipped the script on traditional npcs and created a minecraft agent that mines, crafts, and knows pvp pigs beware … https://t.co/ah7yGj... - @kevskgs | Wed 18, Mar | 18 likes, 6 bookmarks

Takeaway: Promoted to Imagination Chasm and Minecraft Agent. The reviewed video thumbnail shows the stage demo with a Minecraft player view and code visible; the durable point is the demo framing: use a familiar game environment to make an unfamiliar agent workflow imaginable. Source: X/@kevskgs and local video thumbnail review, 2026-07-03

builders can’t craft what they can’t picture, and most industries that agents...

builders can’t craft what they can’t picture, and most industries that agents could transform struggle to define what an "agent" is. imagine if you could plug an agent into every game you played as a kid 🐖 - @itsCathyDi | Fri 13, Mar | 21 likes, 5 bookmarks

Takeaway: Promoted through Imagination Chasm. The durable claim is that industries cannot specify useful agents until builders can picture concrete workflows; the Minecraft example works because it turns "agent" into a familiar game-world action instead of an abstract category. Source: X/@itsCathyDi, 2026-03-13

99% of agents haven’t been built yet

99% of agents haven’t been built yet. models aren't the bottleneck; imagination is. you can't imagine what you can't define. https://t.co/QPSRjtFy5c - @itsCathyDi | Fri 13, Mar | 37 likes, 4 bookmarks

Takeaway: Promoted to Imagination Chasm. The reviewed image shows Cathy speaking on a panel/stage with AI Agent := Model(s) + Tool... visible behind her, which turns the agent definition itself into a concrete artifact. The durable point is the same as the March 17 "useful agents" row: models are not the only bottleneck; builders need definitions and examples before they can picture what to build. Source: X/@itsCathyDi and local image artifact, 2026-07-04

day 7/31 of building Princeton Tower Defense:

day 7/31 of building Princeton Tower Defense: readability day. I overhauled tooltips/modals, made buffs stackable, improved Codex + engineer turret logic, refined mobile selection/hit-testing, updated sprites/previews, and rewrote troop movement indicators + visible healing. https://t.co/yBFYxMww9... - @kevskgs | Thu 19, Mar | 14 likes, 4 bookmarks

Takeaway: Promoted to Princeton Tower Defense. The local image shows the day 7 readability pass as a dense tower-defense state surface: selected-unit detail, stackable buff/healing indicators, tower build tray, spell tray, speed controls, camera controls, and troop overlays. The durable rule is observability before polish; the player needs readable state transitions before prettier UI chrome matters. Source: X/@kevskgs and local image review, 2026-07-04

the next 1 person unicorn isn’t in tech

the next 1 person unicorn isn’t in tech. it’s someone like this. in an industry that hasn’t even defined what an “agent” is yet. - @itsCathyDi | Thu 19, Mar | 18 likes, 4 bookmarks

Takeaway: Promoted through Imagination Chasm. Cathy frames the agent opportunity as vertical and operator-led: the next one-person company may come from a domain that can name the workflow before generic tech teams know what an agent should do there. Source: X/@itsCathyDi, 2026-03-19

Gbrain

@chandangalani @m_jovovich Ah yeah check out my skillpack it does this https://t.co/o5ZPprCPrq - @garrytan | Fri 10, Apr | 2 likes, 4 bookmarks

Takeaway: Search gives you raw pages. GBrain gives you the answer. It's the brain layer your AI agent has been missing — the only one that does synthesis, graph traversal, and gap analysis in one box. Thread: I've been legit impressed by @m_jovovich and MemPalace.

The awesome thing about open source is you can learn a lot about how people do things, an See gbrain_skillpack.md. Source: X/@garrytan, Fri 10, Apr

building an agentic future, one demo at a time

building an agentic future, one demo at a time - @kevskgs | Fri 13, Mar | 12 likes, 3 bookmarks

Takeaway: Promoted into Imagination Chasm as the compact demo thesis: agent futures become believable one concrete demo at a time. No links, thread, replies, media, or versioned artifact were present. Source: X/@kevskgs, reviewed 2026-07-04

there are million/billion dollar valuation startups that just implement agent...

there are million/billion dollar valuation startups that just implement agents in niche industries now imagine being able to bootstrap those agents in seconds… - @kevskgs | Tue 17, Mar | 10 likes, 3 bookmarks

Takeaway: Promoted into Imagination Chasm as the vertical-agent market version of the same rule: niche agent opportunities need concrete before/after workflows before a platform thesis becomes legible. Source: X/@kevskgs, reviewed 2026-07-04

too many subagents in the kitchen

too many subagents in the kitchen - @kevskgs | Wed 18, Mar | 10 likes, 3 bookmarks

Takeaway: Promoted into Multi-Agent Topologies: subagents help only when ownership, merge policy, budgets, cancellation, and verification are explicit. Otherwise the topology adds coordination load. Source: X/@kevskgs, reviewed 2026-07-04

.md files don't make an agent but when agents start making the .md files we'll know agi is here - @kevskgs | Thu 19, Mar | 12 likes, 1 bookmarks

Takeaway: Promoted into Agent Harness: markdown files are harness substrate, not agency. Agency appears when agents read, write, update, act through, and verify those files as durable state. Source: X/@kevskgs, reviewed 2026-07-04

litellm is "soc 2 secured by delve" 🥸

litellm is "soc 2 secured by delve" 🥸 https://t.co/Viu4ObvuZ4 - @kevskgs | Tue 24, Mar | 10 likes, 1 bookmarks

Takeaway: Hub-only compliance-surface artifact. The local screenshot shows LiteLLM's footer with "SOC 2 Type I Secured by Delve" and "ISO 27001 Secured by Delve" badges plus docs links for Gateway, supported providers, Langfuse/OpenTelemetry logging, Prometheus metrics, and virtual keys. Keep it as a searchable trust/compliance signal, not a dedicated LiteLLM capability update. Source: X/@kevskgs and local image review, 2026-07-04

@cursor_ai your agent (using opus 4

@cursor_ai your agent (using opus 4.6 @AnthropicAI) wrote an unhinged 3,400 line response that included: - a mid-meltdown haiku - a fake obituary for itself - a UN Security Council vote - at least 10 different languages - 64 apologies and ended with "Actually, I don't need https://t.co/wVpehgld3E - @kevskgs | Thu 02, Apr | 11 likes, 1 bookmarks

Takeaway: Promoted to Agent Legibility as a recovery-boundary failure. The local media review shows a Cursor-style agent noticing it has no clear question, repeatedly claiming it should stop, then continuing with more finality markers and self-repair text. Durable rule: a harness needs external output budgets, truncation, and summarize/ask gates because model self-talk is not a reliable stop condition. Source: X/@kevskgs and local media review, 2026-07-04

Batch 151–512 — Dedicated-Page Cross-Ref (19 entries)

GBrain v0

GBrain v0.10.0 is a big one My personal OpenClaw setup and brain can now be yours. I've perfected my RESOLVER.md, my SOUL.md and ACLs for multi-user brain access. Now there are 24 distinct fat skills with fat code, fully tested with e2e tests, evals and unit tests. https://t.co/fWe85Hfdey - @garrytan | Wed 15, Apr | 1,673 likes, 2,562 bookmarks

Takeaway: Promoted into Knowledge Brain System and Knowledge System Architecture Comparison. The v0.10 artifact is historical evidence for signal detection, brain-first lookup, media/meeting ingestion, citation repair, task/crons/webhooks, ACLs, and conformant fat skills. Current upstream has moved to 0.42.53.0 with 53 top-level skills, so the 24-skill claim should not be used as current capability coverage. Source: X/@garrytan, Wed 15, Apr; Source: GitHub garrytan/gbrain, 2026-06-30

Introducing Extend UI — open-source components for document agents

Introducing Extend UI — open-source components for document agents - 14 components & examples for PDF, DOCX, and XLSX viewers, plus bounding box citations, file upload, e-signature, and more - fully customizable - MIT licensed when we started, we tried every file viewer and https://t.co/ZA9sEA... - @kushalbyatnal | Tue 09, Jun | 1,503 likes, 2,515 bookmarks

Takeaway: React PDF viewer component with bounding box citation support, page controls, zoom, search, text selection, thumbnails, uploads, and responsive document layouts Source: X/@kushalbyatnal, Tue 09, Jun

Introducing /orchestrate, a skill that recursively spawns agents to tackle yo...

Introducing /orchestrate, a skill that recursively spawns agents to tackle your most ambitious tasks with the Cursor SDK. We’ve used it to: - Autoresearch our internal skills, cutting token use by 20% while improving evals - Cut cold start times on our internal backend by 80% https://t.co/8Hm8S9J5p... - @cursor_ai | Thu 07, May | 2,765 likes, 1,515 bookmarks

Takeaway: This builds on our existing research on multi-agent systems, with a key addition: verifiers.

Planners spawn workers that write code and verifiers that run it. If verification fails, the planner spawns a new worker to fix it. See Cursor Orchestrate Skill. Source: X/@cursor_ai, Thu 07, May

Deep-review note: this row is now promoted into Browser Testing Skills and Harness Engineering as planner/worker/verifier/retry, not just "recursive agents." Source: artifact review, 2026-06-30

Introducing Zero

Introducing Zero The programming language for agents. I wanted a systems language that was faster, smaller, and easier for agents to use and repair. Explicit capabilities. JSON diagnostics. Typed safe fixes. Made for agents on day zero. https://t.co/uTrDOmyBR1 - @ctatedev | Fri 15, May | 2,230 likes, 1,393 bookmarks

Takeaway: Star it, fork it, join the experiment Source: X/@ctatedev, Fri 15, May

Now launching GBrain v0

Now launching GBrain v0.11 with Minions I got sick of OpenClaw's subagents timing out and not getting things done So I built a queue/jobs system that uses GBrain's Postgres/PGLite based on BullMQ to give your OpenClaw/GBrain setup wings. Minions are 10X faster, more reliable https://t.co/1hsgR3I... - @garrytan | Sat 18, Apr | 1,022 likes, 1,339 bookmarks

Takeaway: Promoted to Knowledge Brain System and the 2026-07-02 deep-review section above. Minions is a durable job queue for sub-agent work, not just a launch note. Source: X/@garrytan and local artifact review, 2026-07-02

the web has been quiet for a while

the web has been quiet for a while. for the past few months i've been building something to fix that. declarative audio for the web. describe a sound as plain data, play it with one call. → https://t.co/7NjMJFvFqh https://t.co/1kw7F3EBJ3 - @raphaelsalaja | Tue 21, Apr | 1,391 likes, 1,243 bookmarks

Takeaway: it also comes with its own cli.

browse community made sound packs, install them with one command, or scaffold your own from scratch. Source: X/@raphaelsalaja, Tue 21, Apr

Karpathy's Confusion Protocol is now in GStack

Karpathy's Confusion Protocol is now in GStack Karpathy called it: the #1 AI coding failure mode is the agent confidently picking the wrong path at an ambiguous decision point. You lose 10 minutes of work and have to start over. gstack now has an ambiguity gate built into every https://t.co/6HkrDX... - @garrytan | Thu 16, Apr | 800 likes, 834 bookmarks

Takeaway: Karpathy's Confusion Protocol is now in GStack

Karpathy called it: the #1 AI coding failure mode is the agent confidently picking the wrong path at an ambiguous decision point. You lose 10 minutes of work and have to start over.

gstack now has an a See Andrej Karpathy's LLM Coding Skills. Source: X/@garrytan, Thu 16, Apr

Nowork

AI agent automates Google Ads and SEO https://t.co/57wdHINFd8 https://t.co/UECkxJ9xSD - @tom_doerr | Sat 11, Apr | 385 likes, 653 bookmarks

Takeaway: See NotFair; deep review preserved /toprank:* screenshot drift and refreshed current /notfair:* plugin 0.25.6, Search Console MCP, and notfair-cmo status. Source: X/@tom_doerr, Sat 11, Apr

Just shipped in GStack today: our skills now have greatly upgraded user exper...

Just shipped in GStack today: our skills now have greatly upgraded user experience (UX) capabilities augmented by Steve Krug's principles from "Don't Make Me Think" https://t.co/lbs5XCtlM5 - @garrytan | Tue 14, Apr | 577 likes, 558 bookmarks

Takeaway: Promoted to Quality Review Stack and the 2026-07-02 deep-review section above. The durable content is the UX foundation set: Steve Krug-style behavior principles, six usability tests, first-person narration, $B ux-audit, heatmap screenshots, and token ceiling enforcement. Source: X/@garrytan and local artifact review, 2026-07-02

New Skill: Stripe Emulator

New Skill: Stripe Emulator Works offline ✓ No account needed ✓ Stateful ✓ Hosted checkout page ✓ Webhooks ✓ Embeds in your app ✓ Seedable/deterministic ✓ Works in CI with no secrets ✓ npx skills add vercel-labs/emulate --skill stripe https://t.co/1T9ZDmmJrW - @ctatedev | Wed 29, Apr | 620 likes, 489 bookmarks

Takeaway: Promoted through Emulate (Vercel-labs). Stripe Emulator is a provider-specific skill surface on vercel-labs/emulate, not a new local skill by default: use it when a project needs deterministic checkout/webhook tests with no Stripe account, no network, and no CI secrets. Reviewed frames show a complete local stripe-checkout.emulate.localhost store flow from product page to green Stripe Emulator payment form to paid order confirmation. Source: X/@ctatedev, 2026-04-29; Source: local contact sheet /tmp/xreview-frames/2049350311106511178.jpg, reviewed 2026-07-01

I'm open sourcing it MIT license so we can all speed up and have our own pers...

I'm open sourcing it MIT license so we can all speed up and have our own personal mini-AGI. It's been amazing for me and I want you to have it. To install GBrain in your OpenClaw, just paste this image to your OpenClaw or paste this text: Set up gbrain https://t.co/gFlVvYKifB - @garrytan | Thu 09, Apr | 151 likes, 346 bookmarks

Takeaway: Promoted into Knowledge Brain System. The launch artifact shows the original operator path: install Bun, add github:garrytan/gbrain, run gbrain init --supabase, import an existing markdown brain, query it, read the schema and skillpack docs, then set a daily update check that notifies on new features but never auto-installs. Current source snapshot: 0.42.56.0, MIT, default branch master at 6461790, 24,902 stars and 3,589 forks as of 2026-07-03. Source: X/@garrytan, Thu 09, Apr; Source: GitHub garrytan/gbrain, 2026-07-03

GBrain 0

GBrain 0.9.0 just dropped. Just ask your Claw/Hermes to upgrade if you're already on. Every change I make to my Claw, you get too. And your Claw works with you to pick up the things you want, and custom configure my concepts to what you need. It's the dawn of Personal AI https://t.co/QExmX8qVdn - @garrytan | Sun 12, Apr | 330 likes, 302 bookmarks

Takeaway: Promoted to Knowledge Brain System. The reviewed release-note artifact frames GBrain v0.9.0 as code and skills working together: skills encode judgment calls while deterministic commands handle repeatable work such as publish, backlink repair, lint/fix, reports, and raw-file upload. Current upstream source remains the authority for package version and repo state; the screenshot is historical design evidence for Kevin's scripts-plus-skills operating model. Source: X/@garrytan and local image review, 2026-07-04; Source: GitHub garrytan/gbrain, 2026-07-04

agent-browser v0

agent-browser v0.27 Big day for agents and browsers → React introspection: react tree, react inspect, react renders, react suspense for component trees, props/hooks/state, render profiling, and Suspense analysis → Web Vitals: vitals command reports LCP, CLS, TTFB, FCP, INP + - @ctatedev | Thu 07, May | 461 likes, 278 bookmarks

Takeaway: Promoted into Browser Testing Skills and Chris Tate. v0.27 is the historical release that made agent-browser a React-aware frontend verification tool: react tree, react inspect, render profiling, Suspense analysis, Web Vitals, and batch commands. Current source authority is still agent-browser@0.31.1, with main/v0.31.1 at ed2e10598c9064aecfaeb7cf21b540684db4be2c; the v0.27 tweet is feature provenance, not current version authority. Source: X/@ctatedev and self-thread, 2026-05-07; Source: GitHub/npm, 2026-07-04

Two things I'm excited to ship for GBrain for the next 36 hours:

Two things I'm excited to ship for GBrain for the next 36 hours: Connectors to Claude Cowork / Perplexity Computer so your GBrain can talk with the locked down agents via MCP GBrain Skillpack that will let you have Call-Your-Claw-By-Phone via Twilio with ONE slash command - @garrytan | Sat 11, Apr | 292 likes, 118 bookmarks

Takeaway: Source-reviewed into Knowledge Brain System and Garry Tan. The root post is roadmap/release evidence for GBrain bridging locked-down agents through MCP connectors and a Twilio call-your-agent skillpack; the self-thread says the setup was open source and shipped seven hours later. Current repo state, not the tweet, remains implementation authority. Source: X/@garrytan and self-thread, 2026-04-11; Source: GitHub garrytan/gbrain, checked 2026-07-04

Big wave of security fixes for GStack and GBrain today

Big wave of security fixes for GStack and GBrain today Open Source is incredible. Big thanks to the contributors doing God's work https://t.co/jz7vFjiL50 - @garrytan | Mon 13, Apr | 330 likes, 117 bookmarks

Takeaway: Promoted to Quality Review Stack and Garry Tan. Local image review shows a concrete security-hardening wave for agent browser/skill runtimes: cookie-origin pinning, browse audit logs, symlink/path hardening, setup-script injection fixes, credential redaction, prompt-injection checks, metadata blocking, session-file permissions, frozen lockfile install, and temp-path cleanup. Source: X/@garrytan and local image review, 2026-07-04

new in agent-browser: doctor

new in agent-browser: doctor one command diagnoses problems --fix repairs https://t.co/8PxiVNSLee - @ctatedev | Thu 16, Apr | 196 likes, 87 bookmarks

Takeaway: Promoted to Browser Testing Skills and Chris Tate. Local image review shows agent-browser doctor checking environment, Chrome, daemons, config, providers, network, and launch test; it warns while cleaning one stale daemon, then passes launch in 2.2s. Current source authority remains agent-browser@0.31.1, v0.31.1, repo HEAD ed2e10598c9064aecfaeb7cf21b540684db4be2c. Source: X/@ctatedev and local image review, 2026-07-04

@garrytan 2025 MCPs were supposed to be the next big thing and failed, probab...

@garrytan 2025 MCPs were supposed to be the next big thing and failed, probably because of how much context they were eating up. BUT. This new 2026 MCP architecture with Code Mode is probably that new, done right, magical approach. Just 2 tools that consume a fixed amount of tokens - @anakin | Sat 04, Apr | 9 likes, 13 bookmarks

Takeaway: Promoted through NIA Docs - Docs as Filesystem for Agents as evidence for the filesystem/fixed-tool thesis: many MCP tools can burn context on schemas and descriptions, while Code Mode-style access keeps the tool surface fixed and lets the agent inspect docs/files on demand. No new page was needed because the dedicated NIA docs page already owns the MCP/RAG/filesystem comparison. Source: X/@anakin, 2026-04-04

built contentbit: open source schema-validated content blocks for Markdown

built contentbit: open source schema-validated content blocks for Markdown. LLMs write content that looks done but breaks at the edges: fake components, drifting props, almost-valid tables. contentbit validates with file:line:col errors and gives the model hints to fix itself. https://t.co/hV4eTr9... - @agonist42 | Thu 11, Jun | 20 likes, 10 bookmarks

Takeaway: Contentbit is an agent-output validation loop for Markdown. The deep media/source review added the actual pattern: Markdown is the authoring protocol, the registry/schema is the contract, and the CLI diagnostics give the model line/column hints until output is clean. Current package snapshot: contentbit, @contentbit/core, and @contentbit/react are 0.5.0. Source: X/@agonist42, 2026-06-11; Source: npm registry, 2026-06-30

Releasing Loop

Releasing Loop! An open-source operator desk for agent skills. Loop tracks the sources your skills depend on. When docs change or APIs version, Loop runs a research agent and auto-refreshes your playbooks. Every update is mutable and versioned with a full diff. No silent https://t.co/IBs88ThH1i - @kevskgs | Tue 28, Apr | 11 likes, 4 bookmarks

Takeaway: The catalog already has 100+ verified skills (prompt engineering, MCP development, Next.js patterns, SEO/GEO, security) all refreshing daily and sitting at v16-v21. Been running it on my own skills for a while.

Deep review promotes the launch artifact set to Loop: screenshots show the "Skills that never go stale" hero, a catalog with 111 skills and 45 tracked, the skill-detail/raw-endpoint surface, and a refresh run log with 139.5s, 11 sources, 94 signals, 20 steps, and a revise_skill diff. Current repo/site snapshot: Kevin-Liu-01/Loop, MIT, HEAD c5a62a5, 13 stars, 2 forks, no tags, loooop.dev HTTP 200. Source: X/@kevskgs, Tue 28, Apr; Source: GitHub/site/local image review, 2026-07-03

Ponytail lazy senior dev mode

SOMEONE GAVE CLAUDE CODE A "LAZY SENIOR DEV" MODE AND IT STARTED WRITING 6X LESS CODE. - @om_patel5 | 2026-06-13 | 34 bookmarks

Takeaway: Ponytail is a reference-only code-taste skill/plugin for making agents prove they need new code before writing it. The durable ladder is: skip if unnecessary, reuse existing code, use stdlib, use native platform, use installed dependency, keep one-liners one-line, then write the minimum. Keep it out of the active skill stack until Brin/skill-auditor clear the caution scan; fold the ladder into Code Taste and Security and Review Skills instead. Current source snapshot: MIT, HEAD 40e50d9, latest tag v4.8.4, 72,298 stars, 3,753 forks. Source: X/@om_patel5 and GitHub, 2026-07-03

2026-06-28 Coverage Catch-Up (33 parked records)

---

Timeline

• 2026-06-17 | Marked as a longform compiled bookmark hub and added explicit source provenance. Source: raw/x-bookmarks/bookmarks.jsonl
• 2026-06-13 | Re-synced new AI agent bookmarks and split high-signal items into dedicated pages. Source: raw/x-bookmarks/bookmarks.jsonl
• 2026-05-17 | Created compiled AI/agent bookmark hub from X bookmark batches. Source: raw/x-bookmarks/bookmarks.jsonl