Hermes Agent (Nous Research)

An open-source, self-improving agent runtime ("the agent that grows with you") from Nous Research. It ships the raw primitives of a quality system — self-written skills (procedural memory), persistent cross-session memory, built-in cron to 20+ channels, Slack/Telegram approval buttons, /goal, and multi-agent kanban — that you assemble once and own. It is the runtime Kevin deploys via Agent Machines and the place the The Eval Loop (Slop Is an Output Problem) runs.

What it is

Hermes Agent is Nous Research's open-source agent. It is not on PyPI; install from git: uv pip install 'hermes-agent[web,mcp] @ git+https://github.com/NousResearch/hermes-agent.git@main' (the [mcp] extra is required for mcp_servers to load). Source: Agent Machines constraints, 2026-06-01

There is no "evals" button or quality dashboard. Instead it gives you composable primitives — which is the point: you assemble an eval loop out of pieces the agent already owns, rather than waiting for a feature. Source: User essay "How To Fix AI Slop", 2026-06-01

Current Version Snapshot

As of 2026-07-06, the latest source check found the repository active on main (94cdd56b8263b1f50b962907921ce970549555c7), MIT licensed, and the latest public release still Hermes Agent v0.18.0 (v2026.7.1, published 2026-07-01, tag 462c8b0215c2db3a33252f7456f80a5c4f8acd2b, peeled commit 7c1a029553d87c43ecff8a3821336bc95872213b). The release is the "Judgment Release": P0/P1 closure, evidence-backed completion contracts, first-class MoA model selection, /learn, /journey, background fan-out, and desktop project/coding surfaces. Source: GitHub NousResearch/hermes-agent, 2026-07-06; release v2026.7.1, 2026-07-01

The previous Reach Release widened Hermes from "persistent personal agent" into a multi-surface runtime: iMessage via Photon, Raft network bridge, richer Desktop, background subagents, image editing, Automation Blueprints, Profile Builder, Skills Hub browsing with security scan, atomic memory edits, secure dashboard auth, official WhatsApp Business Cloud API, Telegram rich text, and curator cost optimization. For Kevin's stack, the most important changes were background subagents, safer skill/memory operations, profile-building, and more messaging reach. Source: Hermes release notes, 2026-06-19

Config Defaults and Token Budget

The reviewed config bookmark is useful because it frames Hermes setup as a cost/latency control surface, not just a feature checklist. Official docs confirm the most explicit knob: agent.reasoning_effort defaults to medium when unset, accepts none, minimal, low, medium, high, and xhigh, and trades better performance on complex tasks for higher token use and latency. Use lower effort for routine cron/profile tasks and reserve high or xhigh for hard architecture, research, or review work that benefits from deeper thinking. Source: X/@IBuzovskyi and local video contact sheet, 2026-07-04; Source: Hermes configuration docs, 2026-07-04

The local video artifact is mostly product/setup evidence: Hermes branding, setup/status cards, Slack confirmation, and a settings/config grid. It does not independently reveal all eight claimed config values, so the current wiki claim stays narrow: the bookmark deep-review promotes the token-budget/config-default habit and the docs-backed reasoning_effort setting, not an unsourced eight-item checklist.

Community Reference: Hermes Bible

Hermes Bible is an unofficial search and workflow index over Hermes docs and community material. The live site advertises 169 pages and 25 flows and explicitly points readers back to the official Hermes documentation as the source of truth. Use it for discovery and example-finding; use this page, GitHub releases, and Nous docs for source authority. Source: Hermes Bible site, 2026-07-04

Creative App Computer-Use Loop

Amber Shen's TouchDesigner demo is useful because it shows Hermes as a creative-app operator, not only a coding/runtime shell. The reviewed video frames show a polished launch/demo sequence, Hermes desktop/provider settings, and a Minimax M3 / VLM setup frame; the tweet's workflow says the agent navigated the desktop, discovered how to connect to TouchDesigner, read reference images, iterated with the user, and saved the output. Route this as a Computer Use and Browser Automation Patterns and creative-tool learning loop: the agent can acquire a GUI workflow from docs, screenshots, reference images, and user critique, but the durable output still needs artifact evidence and human taste review before it becomes a reusable skill. Source: X/@whosamberella, 2026-06-13; Source: local video contact sheet, 2026-07-02

LLM-Wiki Bundled Skill

Hermes packages a bundled llm-wiki research skill for building and querying interlinked Markdown knowledge bases. The docs describe it as Karpathy's LLM Wiki pattern: compile knowledge once into Markdown pages, preserve cross-references, track contradictions, and use the agent for summarizing, filing, and consistency maintenance. Current docs list the bundled skill at version 2.1.0 under skills/research/llm-wiki, with WIKI_PATH configurable and ~/wiki as the default. Source: Hermes LLM-Wiki docs, reviewed 2026-07-02

The reviewed image shows the practical output: a dark graph view of a Nous research wiki with dense nodes around grpo, rl-post-training, index, nemo-rl-vs-alternatives, atropos, and nous-research, plus outlier source pages and schema/wikilink nodes. Teknium's self-thread called out GRPO as the most interconnected topic. The durable signal is not only that Hermes can write notes; it can emit a graph-shaped research vault from web/code/paper study, making Karpathy LLM Wiki Setup Guide a first-class agent workflow rather than an external note-taking habit. Source: X/@Teknium thread, 2026-04-07

Blank Slate Setup Mode

Hermes' Quickstart now offers Blank Slate as a first-run setup path. It starts with only provider/model selection, the File Operations toolset, and the Terminal toolset. Web, browser, code execution, vision, memory, delegation, cron, skills, plugins, MCP servers, compression, checkpoints, smart routing, and memory capture start disabled unless the operator opts in. The generated config explicitly records platform_toolsets.cli and agent.disabled_toolsets, so later hermes update runs do not silently load tools the user never chose. Source: Hermes Quickstart docs, reviewed 2026-06-30

The bookmark media is mostly branding (a "Blank Slate Mode" thumbnail and video title treatment), so the durable signal is the docs-backed setup contract. For Kevin, this adds a clean profile pattern: start any new Hermes or agent-machine profile from the minimum trusted tool surface, then add web, memory, skills, cron, delegation, and MCP only when the task actually needs them. Source: X/@NousResearch, 2026-06-20

Profile Distributions

Profile distributions turn a Hermes profile into a shareable git-backed agent package: distribution.yaml, SOUL.md, config, skills, MCP connections, cron jobs, files, and other runtime setup travel together, while the receiver's memories, sessions, API keys, and local secrets stay untouched. The docs frame this as the difference between sending someone a pile of setup notes and giving them a versioned agent they can install and update in place. Source: Hermes profile distributions docs, 2026-06-30

This is the cleanest Hermes analogue to Kevin's Agent Harness Capsules: expose the public brain and operating surface as a package, but keep private state outside the distribution boundary. Teknium's thread also clarifies the default-skill philosophy: out-of-box profiles should be opinionated enough to save weeks of setup, while optional skills absorb niche breadth as the catalog grows. Source: X/@Teknium thread, 2026-05-31

Primitives (the pieces you assemble)

• Skills = procedural memory — the agent writes, keeps, and reuses skills; they compound. Nous reports agents with 20+ self-created skills finish similar tasks ~40% faster because they stop rediscovering the process. Encode taste once (e.g. an LLM-as-a-judge rubric) and it grades forever. Source: User essay, 2026-06-01
• Persistent memory — grows across sessions with full cross-session recall; your ground truth (e.g. gold-standard examples) lives here, queryable.
• Built-in cron — scheduled jobs that deliver to 20+ channels; run a sampler over production and DM you the moment a score dips.
• Approval buttons — native into Slack and Telegram, so the agent works in the background and taps you when a decision is yours ("scores went 0.81 → 0.74, two cases regressed, approve?").
• /goal — holds the agent to a target across turns (the completion gate).
• Multi-agent kanban — decompose a run, score in parallel, schedule it.
• Self-improvement core — not a bolted-on feature; flagged failures get written back into skills as new checks, so the system hardens on its own. Source: User essay, 2026-06-01

Teknium's Open Source AI endorsement is best read as a stack-ownership signal for this primitive set, not a separate release note: Hermes is useful because the skills, memory, cron, evals, and self-improvement loop can live in infrastructure the operator owns. For current feature/version authority, keep using the GitHub release and official docs snapshot above. Source: X/@Teknium, 2026-06-15

Kanban Block Loop Breaker

Teknium's Kanban update makes the self-improvement loop more explicit at the task layer. The reviewed diagram names the old failure mode: block -> cron unblocks -> re-block -> repeat forever. The fix is typed block reasons (dependency, needs_input, capability, transient) plus a block_recurrences counter that survives unblocks and resets only when the task is done. Dependency blocks become TODOs that auto-resume when their parents finish; repeated same-cause re-blocks route to a human instead of looping. Source: X/@Teknium and local artifact review, 2026-07-03

What you build with it

The canonical application is the The Eval Loop (Slop Is an Output Problem) — a quality gate for AI output. The mapping: skills → the judge + the suite; memory → the benchmark's ground truth; cron → production monitoring; approval buttons → the ship gate; self-improvement → failures becoming new test cases. See The Eval Loop (Slop Is an Output Problem) for the six-move build (stand up + reachable, load gold standard, judge skill, suite skill, gate with regression + approval, watch production on cron).

/learn Skill Creation (Jun 2026)

Hermes added /learn: pass it sources, past sessions, URLs, docs, directories, PDFs, configs, or manuals, and Hermes distills a verifiable reusable skill from the material. The launch framing is important: the agent does not merely summarize the sources; it builds the skill, tests it live, and crystallizes the new procedure for reuse. Source: X/@NousResearch, 2026-06-24; Source: X/@Teknium, 2026-06-24

The implication for Kevin's wiki is direct: recurring work should still follow Browser Testing Skills and the no-one-off lifecycle, but /learn is now a candidate accelerator for the prototype → codify step. Feed it the source corpus and past successful transcripts; require it to produce a tested SKILL.md, not just a draft prompt. Akshay's visual makes the source surface concrete: local SDK directories, docs URLs, workflows just walked through, and pasted notes can feed one normal agent turn; the saved skill should be reusable before the task repeats. Source: X/@akshay_pachaar and local image review, 2026-07-04

The @jumperz visual adds the missing operator pattern: do not dump random memory into /learn. Stage raw material in a private "Learnables" room, split each topic into a thread, let Hermes gather docs/examples/pitfalls, add the operator's taste as explicit keep/cut rules, then say "learn this" only after the thread is coherent. The output should be saved as a reusable artifact or skill and should not overwrite prior taste by default. That makes /learn a codification lane inside Kevin's no-one-off lifecycle rather than a new flavor of note hoarding. Source: X/@jumperz and local image review, 2026-07-04

Implementation detail: /learn is not a separate ingestion service. Current Hermes docs describe it as a standards-guided prompt that lets the live agent gather the material, write a SKILL.md, and save it with skill_manage; if skills.write_approval is enabled, the normal write-approval gate applies. The dashboard exposes the same flow through a "Learn a skill" panel with directory, URL, and free-text inputs. Source: Hermes skills docs, 2026-06-30

Local demo media: thumbnail; video remains available in the original tweet/enriched source record.

Mixture of Agents 2.0

Hermes now exposes Mixture of Agents as a virtual model provider: each named MoA preset appears under the moa provider, reference models run first, and the aggregator becomes the acting model that writes the final response and emits tool calls. The important part is that MoA stays inside the normal Hermes loop: session context, tools, interrupts, transcript persistence, and surfaces such as CLI/gateway/TUI keep working. Source: Hermes docs, 2026-06-27

The practical reading for Kevin is model-council-in-runtime, not a separate research trick. Use MoA when a hard task benefits from multiple model perspectives but must still obey Hermes skills, memory, approval, and tool-use behavior. Teknium's launch note also says Nous is benchmarking cheaper open-source model combinations against Opus/GPT-5.5-style baselines, so route through The Eval Loop (Slop Is an Output Problem) before accepting a preset as cheaper or better. Source: X/@Teknium, 2026-06-26

Artifact-review note: one reply in this thread pointed to PR #37272, but GitHub shows that PR as open and titled feat(skills): add dynamic-workflow orchestration skill. The MoA docs are the feature authority for this page; PR #37272 is a related skill-orchestration artifact, not proof that MoA itself shipped through that PR. Source: GitHub PR NousResearch/hermes-agent#37272, 2026-06-30

Memory Provider and Context Engine Plugins

Hermes exposes memory and context management as single-select provider surfaces, which is more important than another memory product name. A memory provider plugin adds persistent cross-session knowledge beyond built-in MEMORY.md / USER.md; a context engine plugin can replace the built-in compressor with another context strategy, such as a knowledge DAG. Both are config-driven and managed through hermes plugins, and both enforce one active provider/engine at a time to avoid tool-schema bloat and conflicting backends. Source: Hermes memory provider docs, 2026-06-30; Source: Hermes context engine docs, 2026-06-30

The useful pattern for Kevin is "retrieval should live in the configurable memory slot when it is used constantly." mr-r0b0t's screenshot shows /learn producing a hermes-capability-reference skill from the Hermes developer docs, but their reply thread notes the remaining friction: if the agent repeatedly invokes a docs skill just to know what Hermes can do, that knowledge may belong in the memory provider or context engine layer instead. Source: X/@mr_r0b0t thread, 2026-06-26; Source: r0b0tlab LLM-wiki repo

Kevin's current operator-stack decision is to use Hindsight Memory Provider as the active Hermes long-memory provider and Honcho as the profile/user-modeling layer through MCP/SDK, a separate profile, or an explicit bridge. The reason is structural: Hermes memory providers are single-select, while Kevin's Kevin-Wiki plus qmd remains the canonical source of truth. Runtime memory can recall and reflect; durable knowledge still writes back to wiki pages, skills, automations, config, and logs. Source: User request, 2026-07-04; Hermes memory providers docs

The current user guide lists eight external provider plugins—OpenViking, Honcho, Mem0, Hindsight, Holographic, RetainDB, ByteRover, and Supermemory. Only one external provider can be active at once, but built-in MEMORY.md and USER.md remain active alongside it. When enabled, Hermes injects provider context, prefetches relevant memories, syncs turns, optionally extracts on session end, mirrors built-in writes, and exposes provider-specific tools. This confirms the provider is an additive runtime recall layer, not a replacement for canonical wiki writeback. Source: Hermes memory providers docs, reviewed 2026-07-14

Custom Plugin Guide (Jun 2026)

The Miles Deutscher guide video is a docs/code walkthrough, not a separate package. The reviewed frames show plugin.yaml, tool registration, plugin discovery, categories, subfolders, gateways, and actions. The durable route is Hermes' official plugin surface: put a directory under ~/.hermes/plugins/my-plugin/ with plugin.yaml, __init__.py, schemas.py, and tools.py when Hermes needs custom tools, hooks, integrations, API keys, binary data, streaming, or project/team-specific runtime behavior. Source: X/@milesdeutscher and local video contact sheet, 2026-07-02; Source: Hermes plugins docs, 2026-07-02

The skill/tool boundary is explicit in current docs: use a skill when the capability is instructions, shell commands, or orchestration over existing tools; use a plugin/tool when Hermes needs new executable logic or a new external integration. Add built-in core tools only when the capability belongs in Hermes itself. This maps cleanly to Kevin's no-one-off rule: prototype a repeated workflow as a skill first; promote to a Hermes plugin only when instructions are insufficient and the runtime needs a new tool surface. Source: Hermes "Build a plugin" guide and "Adding tools" developer guide, 2026-07-02

Unreal Engine MCP Catalog Entry (Jul 2026)

Teknium's artifact is useful because it turns Epic's experimental Unreal Engine 5.8 MCP server into a Hermes catalog-install route. The reviewed card says the plugin is official/embedded in the editor, speaks local HTTP at 127.0.0.1:8000/mcp, has no auth, and can spawn actors, configure lighting, and create materials. Epic's docs confirm the safety boundary: same-machine local HTTP only, no remote/auth layer, and tool calls execute serially on the game thread, so clients should not issue overlapping tool calls. Hermes' MCP docs supply the other half: catalog install/discovery, remote HTTP and local stdio MCP server config, and per-server tool filtering. Treat this as catalog convenience over an experimental local editor bridge, not a production remote-control surface. Source: X/@Teknium and local artifact review, 2026-07-03; Source: Epic Unreal MCP docs, 2026-07-03; Source: Hermes MCP docs, 2026-07-03

Creative Ideation Optional Skill (Jun 2026)

Nous published a Hermes optional skill for creative ideation: the skill analyzes a user's prompt, selects among 22 creative methodologies, and routes the agent through structured ideation instead of generic brainstorming. The operational lesson is broader than Hermes: creative procedures belong in reusable skills, with explicit method selection and follow-up prompts. Source: X/@NousResearch, 2026-06-24; Source: GitHub optional skill

Enable optional skills in Hermes with hermes skills config, then keep the same acceptance bar Kevin uses for local skills: test on real prompts, record what methods were selected, and only promote if the output beats the baseline.

Source status: upstream optional-skills/creative/creative-ideation/SKILL.md reports version 2.1.0 and depends on a references/ method library. It is a strong candidate for a local Kevin skill, but should not be activated blindly; import the full folder, run skill-auditor, test against Kevin's ideation/taste examples, and then decide whether it belongs under skills/personal/ or remains Hermes-only. Source: GitHub optional skill, 2026-06-30

Dashboard: Profile Builder (Jun 2026)

Nous shipped a Profile Builder in the Hermes dashboard: a wizard flow to define agent identity end-to-end. The local video review shows the same five-step frame as the docs and thread: Identity → Model → Skills → MCPs → Review. Source: X/@NousResearch thread, 2026-06-10; Source: Hermes Web Dashboard docs, 2026-06-30

1. Sidebar → Profiles → Build (next to quick Create)
2. Steps: Identity → Model → Skills → MCPs → Review
3. Controls: name/description, model/provider, built-in + optional skills, skills-hub installs, MCP servers

3,022 likes and 1,452 bookmarks were recorded on the reviewed announcement. This lowers harness setup friction for multi-agent fleets because profiles can be assembled by intent rather than by editing YAML directly. It also mirrors Kevin's capsule boundary: agent identity, model, skills, and MCPs are public/runtime configuration; memories, sessions, and secrets stay outside the portable profile. Source: X/@NousResearch, 2026-06-10

Automation Blueprints (Jun 2026)

Nous shipped Automation Blueprints — turning cron jobs into "clickable, fillable, conversational workflows" (parameterized forms instead of raw cron syntax). Three trigger types, each model-agnostic and deliverable to Telegram/Discord/Slack/SMS/email/GitHub comments/local files: Source: Hermes docs "Automation Blueprints", https://hermes-agent.nousresearch.com/docs/guides/automation-templates, 2026-06-11

The shipped blueprint catalog maps almost 1:1 onto Kevin's own automations/: Nightly Backlog Triage, Automatic PR Code Review, Docs Drift Detection, Dependency Security Audit, Deploy Verification, Alert Triage — each a copy-paste hermes cron create / hermes webhook subscribe recipe with a [SILENT] no-op convention. 2,835 likes and 1,518 bookmarks were recorded on the reviewed announcement. Source: X/@NousResearch, 2026-06-11; Source: Hermes docs, 2026-06-11

The video contact sheet shows the durable product framing: "Schedule anything", "a blueprint for every routine", and "one function for every path." This is the cron primitive productized into a UI: the same engine that runs Kevin's The Eval Loop (Slop Is an Output Problem) sampler now has fillable forms plus GitHub/webhook triggers, directly portable to the Agent Machines deployment.

Source-review note: the docs-only follow-up bookmark (2065200919038292326) resolves to the Automation Blueprints guide and a root announcement video link that Jina blocks. The docs are the authority for trigger taxonomy and delivery channels; the root announcement media remains the visual artifact. As of 2026-07-04 the guide's published timestamp is 2026-07-03 19:39 UTC, and it still lists Schedule, GitHub Event, and API Call as the three trigger families. Source: X/@NousResearch, 2026-06-11; Source: Hermes Automation Blueprints docs, checked 2026-07-04

Hermes Desktop

Same week: Hermes Desktop — native macOS/Windows app wrapping the Hermes experience (@NousResearch, 12.7K likes). Source: X/@NousResearch, 2026-06-09

Browser Extension and Operator Field Notes (Jul 2026)

The July bookmark batch adds a browser-native side panel to the Hermes stack. abundantbeing/hermes-browser-extension is MIT licensed and active at HEAD 07e6858d4d945e2fc2c956686c5db6b9e18b5067, with tags through v0.1.9. The local launch thumbnail is a Hermes Browser Extension brand card; treat it as evidence for a web-context bridge into local or remote Hermes runtime, not as source authority for exact runtime behavior. The durable route is Agent Cookie - Session State Sync for the Agent's Second Mac / Computer Use and Browser Automation Patterns: browser context is useful when the agent needs the active page, but auth/session boundaries stay sensitive. Source: X/@Teknium, 2026-07-05; Source: GitHub abundantbeing/hermes-browser-extension, 2026-07-06

The operator rows are useful as field notes, not as a reproducible setup recipe. They describe a small always-on box and fallback model/provider choices around Codex OAuth, Fable, GLM, and Kimi, plus personal-work surfaces such as notes, messaging, and Linear. Preserve the pattern: Hermes is the persistent runtime and memory/router; coding harnesses remain separate tools it can call. Do not copy the exact account/app list without a security review. Source: X/@_jonatasantos, 2026-07-03; Source: X/@_jonatasantos, 2026-07-04

The privacy/data-broker rows now route to focused page Unbroker. Hermes documents Unbroker as an optional official/security/unbroker skill at version 1.0.0, installable with hermes skills install official/security/unbroker, authored by SHL0MS, MIT licensed, and aimed at finding/removing exposed personal information from broker sites. It can be valuable, but it touches identity data and third-party compliance workflows, so Kevin's route is approval-gated: recorded consent, reviewed disclosure fields, audit logging, narrow credential scope, and no unattended CAPTCHA/ID/phone/fax work. The same caution applies to product claims about web extraction being "60x faster" or "49x cheaper": record the signal, but require release notes, benchmarks, and a task-matched eval before using it as an operational cost assumption. Source: X/@Teknium, 2026-07-03/04; Source: Hermes Unbroker docs, 2026-07-06; Source: X/@NousResearch, 2026-07-01

The rest of the cluster reinforces existing page decisions: v0.18/Judgment remains the release authority, sovereignty/open-source framing belongs under Own the Learning Loop, and browser/dashboard/operator claims must be split from docs-backed feature claims. Source: X bookmark artifact audit, 2026-07-06; rows 2072416088785359189, 2073201424054562859, 2073041729901473932, 2071938033440763985

Operator discourse (bookmark batch)

• @Teknium: Hermes is not a blank-slate agent — opinionated defaults + skills out of the box. Source: X/@Teknium, 2026-06-08
• @_jonatasantos stack recipe: VPS + Hermes + Knowledge Brain System or a qmd/SQL memory vault + Codex auth + Orca/Tailscale, with Hermes positioned as the persistent always-on agent and Claude Code/Codex as coding harnesses it can call. His replies make the boundary explicit: Hermes is for shared context, routing, evals, memory, and multiple projects; Claude Code is a harness focused on writing code. Source: X/@_jonatasantos thread, 2026-06-01
• Peter Steinberger's linked "Just Talk To It" article reinforces the operator loop: think in blast radius, stop an agent mid-run when the actual scope diverges, ask for status, then continue/abort/split based on evidence. Kevin keeps Peter's status/blast-radius habit but not the unsafe parts for shared checkouts: parallel agents still need owned commits or isolated worktrees unless a single human is deliberately supervising one tree. Source: Peter Steinberger, read 2026-06-30
• @milesdeutscher: Nous full guide on custom Hermes agent plugins; routed above to the official plugin/tool boundary. Source: X/@milesdeutscher, 2026-06-07

Relationship to Kevin's stack

• Agent Machines is Kevin's persistent Hermes deployment on a Dedalus microVM (bundled skills, crons, Cursor SDK bridge). Hermes is the runtime; Agent Machines is the rig around it. (OpenClaw is the alternate runtime on the same abstraction.)
• Agent Cookie - Session State Sync for the Agent's Second Mac can keep a headless Hermes box authenticated as Kevin across his sites/APIs.
• The skills it runs are synced from kevin-wiki/skills/personal/ — the same taxonomy as this wiki.

---

Timeline

• 2026-07-14 | Refreshed the official memory-provider model: eight external plugins, one active at a time, with built-in memory always additive and Kevin-Wiki/qmd remaining canonical. Source: User request, 2026-07-14; Hermes memory providers docs

• 2026-07-12 | Added Hermes Cloud, Desktop discovery, direct portal deployment, and /goal as product-shape evidence. They strengthen the persistent-agent route but do not replace the local Hermes operator boundary or establish a version claim beyond official docs. Source: X Bookmark Ingest 2026-07-12; X/@NousResearch and @iamlukethedev, 2026-07-08–10

• 2026-07-06 | Added July Hermes field notes: browser extension source at abundantbeing/hermes-browser-extension HEAD 07e6858, operator stack notes, personal-data-removal skill caution, and web-extraction speed/cost claims as benchmark-required product signals. Refreshed NousResearch/hermes-agent main HEAD to 94cdd56b8263b1f50b962907921ce970549555c7; latest public release remains v0.18.0 / v2026.7.1. Source: X bookmark artifact audit, 2026-07-06

• 2026-07-04 | Reviewed Teknium's Open Source AI / Hermes self-improvement endorsement as stack-ownership evidence. No new Hermes version or feature claim was added beyond the current release/docs snapshot. Source: X/@Teknium, 2026-06-15

• 2026-07-04 | Deep-reviewed the @IBuzovskyi Hermes config artifact. Added the docs-backed agent.reasoning_effort token/latency control and preserved the visual review boundary: the video shows Hermes setup/config UI, but the docs are the authority for exact config values. Source: X/@IBuzovskyi, 2026-06-21; Source: Hermes configuration docs, 2026-07-04

• 2026-07-04 | Source-reviewed the docs-only Automation Blueprints follow-up. Confirmed the current guide still uses the three-trigger taxonomy (Schedule, GitHub Event, API Call), any-model execution, and delivery to Telegram/Discord/Slack/SMS/email/GitHub comments/local files; no new local media was expected for the follow-up row. Source: X/@NousResearch, 2026-06-11; Source: Hermes docs, checked 2026-07-04

• 2026-07-04 | Added Hermes Bible as an unofficial community docs/search layer after source review of the Reach Release bookmark. Updated current GitHub source snapshot to main HEAD 5445e42 while latest public release remains v2026.7.1 / 0.18.0. Source: X/@iamlukethedev, 2026-06-19; Source: Hermes Bible, 2026-07-04; Source: GitHub NousResearch/hermes-agent, 2026-07-04

• 2026-07-03 | Deep-reviewed Teknium's Unreal Engine MCP Catalog artifact. Added the local-editor safety boundary and the Hermes MCP catalog route: same-machine HTTP, no auth, serialized game-thread calls, catalog install/discovery, and per-server tool filtering. Source: X/@Teknium, 2026-06-18; Source: Epic Unreal MCP docs, 2026-07-03

• 2026-07-03 | Deep-reviewed Teknium's Hermes Kanban block-loop artifact. Added typed block reasons, dependency auto-resume, human escalation after repeated same-cause blocks, and current source snapshot at HEAD 662426e. Source: X/@Teknium, 2026-06-26

• 2026-07-02 | Deep-reviewed the Miles Deutscher custom-plugin guide video and official Hermes plugin docs. Captured the route: skills for instructions and existing-tool orchestration; plugins for custom tools, hooks, integrations, API keys, binary data, streaming, and runtime behavior. Current source snapshot moved to main HEAD 30e947e0 with 207,728 stars and 37,720 forks. Source: X/@milesdeutscher, 2026-06-07; Source: Hermes plugin docs and GitHub, 2026-07-02

• 2026-07-02 | Deep-reviewed Amber Shen's Hermes + TouchDesigner artifact. Added it as evidence that Hermes can operate creative GUI tools through a self-learning computer-use loop: desktop navigation, tool connection discovery, reference-image reading, iterative critique, and saved output. Source: X/@whosamberella, 2026-06-13

• 2026-07-02 | Deep-reviewed the Teknium Hermes + LLM-Wiki artifact row. Added the current v0.18.0 release snapshot and the bundled llm-wiki skill evidence: version 2.1.0, configurable WIKI_PATH, Markdown/Obsidian-compatible graph structure, and a reviewed local graph image showing a Nous research wiki clustered around GRPO/RL post-training topics. Source: X/@Teknium, 2026-04-07; Source: Hermes LLM-Wiki docs, 2026-07-02

• 2026-06-30 | Deep-reviewed Jonata Santos's Hermes stack recipe and the linked Peter Steinberger article. Promoted the durable boundary: Hermes owns persistent memory/routing/evals/multi-project context, while Claude Code/Codex remain coding harnesses; adopt the blast-radius/status/abort habit without weakening Kevin's shared-worktree git safety. Source: X/@_jonatasantos, 2026-06-01; Source: Peter Steinberger, 2026-06-30

• 2026-06-30 | Deep-reviewed the Profile Builder and Automation Blueprints artifact rows. Added root tweet embeds, local media cards, current GitHub HEAD/version snapshot, and the durable reading: profile creation and automation setup are moving from raw YAML/cron into form-backed workflows that still compile to explicit runtime configuration. Source: X/@NousResearch, 2026-06-10; Source: X/@NousResearch, 2026-06-11

• 2026-06-30 | Added Profile Distributions as the shareable-agent packaging layer: versioned public brain/config/skills/cron/MCP in git, private memory/sessions/secrets outside the package. Source: X/@Teknium, 2026-05-31; Source: Hermes profile distributions docs, 2026-06-30

• 2026-06-30 | Added Blank Slate setup mode from Hermes Quickstart: start with provider/model plus file and terminal tools only, with web, browser, memory, skills, cron, delegation, plugins, MCP, compression, checkpoints, and smart routing disabled until explicitly opted in. Source: X/@NousResearch, 2026-06-20; Source: Hermes Quickstart docs, 2026-06-30

• 2026-06-30 | Deep-reviewed six Hermes-related X bookmark rows: Reach Release v0.17.0, /learn, MoA 2.0, creative-ideation, and memory provider plugins. Verified GitHub release/tag/HEAD state, repaired local media, and promoted the durable capsule rule: skill docs, memory providers, context engines, cron, and model councils are separate harness surfaces. Source: X bookmark artifact audit, 2026-06-30; GitHub NousResearch/hermes-agent, 2026-06-30

• 2026-06-30 | Added direct Hermes Desktop tweet and local media artifact block. Source: X/@NousResearch, 2026-06-09

• 2026-06-28 | Added Mixture of Agents 2.0 as a virtual model provider inside Hermes, with presets selectable across normal Hermes surfaces and benchmark claims requiring eval-loop validation. Source: X/@Teknium, 2026-06-26; Source: Hermes docs

• 2026-06-25 | Captured the Hermes optional creative-ideation skill: prompt analysis routes through 22 creative methodologies, reinforcing skill-based creative workflows over generic brainstorming. Source: X/@NousResearch, 2026-06-24; Source: GitHub optional skill

• 2026-06-25 | Added Akshay's /learn explainer screenshot as reinforcement for the self-improving skill loop: learn from source material before repeating the painful live session. Source: X/@akshay_pachaar, 2026-06-25

• 2026-06-24 | /learn launched: Hermes can ingest source material and synthesize/test a reusable skill. Captured as a skill-creation accelerator for Kevin's no-one-off workflow. Source: X/@NousResearch, 2026-06-24; Source: X/@Teknium, 2026-06-24

• 2026-06-15 | Satya Nadella's "A frontier without an ecosystem is not stable" essay validates the own-it-yourself framing: the firm's moat is an owned, model-portable learning loop (swap the generalist model without losing "company veteran" expertise) — exactly Hermes's self-written skills + persistent memory + self-improvement, assembled once and owned. Kevin: "couldn't have written a better blog… in favor of Hermes Agent's self-improvement loop." See Own the Learning Loop. Source: X/@satyanadella, https://x.com/satyanadella/status/2066182223213293753; User, 2026-06-15

• 2026-06-11 | Automation Blueprints shipped — cron jobs become clickable/fillable/conversational workflows; three trigger types (schedule/GitHub event/API call), catalog mirrors Kevin's own automations (PR review, backlog triage, docs drift, dep audit). 2,733 likes, 1,472 bookmarks. Source: X/@NousResearch, 2026-06-11; Source: Hermes docs, 2026-06-11

• 2026-06-10 | Profile Builder in dashboard — Identity → Model → Skills → MCPs → Review wizard. Source: X/@NousResearch, 2026-06-10

• 2026-06-09 | Hermes Desktop native app announced. Source: X/@NousResearch, 2026-06-09

• 2026-06-01 | Page created from Kevin's "How To Fix AI Slop (Using Hermes)" essay — first dedicated page for the Nous Research runtime (previously only referenced inside Agent Machines). Captured the primitives (skills/memory/cron/approvals//goal/kanban/self-improvement) and the link to the The Eval Loop (Slop Is an Output Problem) it's built to run. Source: User, 2026-06-01