AI SDK HarnessAgent

AI SDK 7 unifies established agent harnesses behind one TypeScript API — swap Claude Code, Codex, Deep Agents, OpenCode, or Pi like you swap models, with sandboxed sessions and AI SDK-compatible streams.

Vercel shipped HarnessAgent on 2026-06-12 as an experimental canary API. The thesis: the AI SDK already made models swappable; harnesses manage everything above a model call (skills, sandboxes, sessions, permissions, compaction, sub-agents, runtime config) — and should be swappable the same way. Source: Vercel changelog, https://vercel.com/changelog/program-agent-harnesses-with-ai-sdk, 2026-06-12

What it is

Two abstractions, decoupled:

A harness is a complete agent runtime — not a thin wrapper around one model call. It owns workspace access, built-in coding tools, native session state, compaction, permission flows, and harness-specific configuration. Every harness run operates inside a sandbox so the host environment stays safe. Source: AI SDK docs, https://ai-sdk.dev/v7/docs/ai-sdk-harnesses/overview, 2026-06-12

HarnessAgent.generate() and .stream() return standard AI SDK result types (GenerateTextResult, StreamTextResult). Apps using useChat or related UI tooling can consume harness streams without rewriting the UI layer — including toUIMessageStream for chat surfaces. Source: Vercel changelog, 2026-06-12

Install

npm i ai @ai-sdk/harness @ai-sdk/sandbox-vercel

Package split (experimental - expect breaking changes):

Docs: AI SDK harnesses overview

Minimal usage

import { HarnessAgent } from '@ai-sdk/harness/agent';
import { claudeCode } from '@ai-sdk/harness-claude-code';
import { createVercelSandbox } from '@ai-sdk/sandbox-vercel';

const agent = new HarnessAgent({
  harness: claudeCode,
  sandbox: createVercelSandbox({ runtime: 'node24', ports: [4000] }),
  tools: { /* custom AI SDK tools */ },
  skills: [ /* custom skills */ ],
});

const session = await agent.createSession();
try {
  const result = await agent.stream({
    session,
    prompt: 'Check the test failures and fix the production code.',
  });
  for await (const part of result.fullStream) {
    if (part.type === 'text-delta') process.stdout.write(part.text);
  }
} finally {
  await session.destroy();
}

Swap claudeCode for codex, deepAgents, openCode, or pi — same HarnessAgent flow. Source: Vercel changelog, 2026-06-12; Vercel changelog, 2026-06-25

Deep Agents And OpenCode

On 2026-06-25, Vercel added two harness adapters:

• @ai-sdk/harness-deepagents adapts LangChain's deepagents runtime, including file and shell tools, skills, host tools, multi-turn sessions, attach/resume, and built-in tool approvals.
• @ai-sdk/harness-opencode boots an OpenCode server inside the sandbox via @opencode-ai/sdk, streams session events through the harness, exposes OpenCode tools, supports built-in and host tool approvals, and lets the app choose model, provider, and reasoning variant.

The supported harness set is now Claude Code, Codex, Deep Agents, OpenCode, and Pi. Current package snapshot as of 2026-07-04: @ai-sdk/harness@1.0.18, @ai-sdk/harness-deepagents@1.0.17, and @ai-sdk/harness-opencode@1.0.18, each on the latest dist-tag; snapshot/canary channels also exist, so check the exact package before coding against adapter config. Source: Vercel changelog, 2026-06-25; Source: npm registry via npm view, 2026-07-04

Core concepts

• HarnessAgent — application-facing agent; accepts custom instructions, skills, AI SDK tools, permission settings, sandbox hooks, adapter-specific config while preserving native harness behavior.
• Harness adapter — connects to a runtime (@ai-sdk/harness-claude-code, etc.).
• Sandbox provider — isolated filesystem + process environment (createVercelSandbox).
• Session — live conversation + workspace state; create before turns, destroy() when done. Server routes can persist resume state via session.detach() / session.stop() with a stable sessionId. Source: AI SDK docs, https://ai-sdk.dev/v7/docs/ai-sdk-harnesses/overview, 2026-06-12

When to use a harness vs a model loop

Use a harness when an existing runtime should drive the task:

• Coding agents inspecting/modifying a sandboxed workspace
• Runtimes with built-in tools and permission models (Claude Code Harness, Codex Harness (OpenAI))
• Multi-turn sessions where the runtime owns conversation history and compaction
• Workflows that should preserve native harness behavior instead of re-implementing a tool loop

Use providers/models when you need direct control over the tool loop, model settings, structured output, or a custom agent architecture (AI SDK 5 Agent, stopWhen, prepareStep). Source: AI SDK docs, 2026-06-12

Stream compatibility

Harness events project into AI SDK stream parts where possible: text, reasoning, tool calls/results, usage, finish reasons. Events without first-class AI SDK parts (workspace file changes, compaction) surface as dynamic provider-executed tool parts. Source: AI SDK docs, 2026-06-12

Stack fit

• Parent toolkit: Vercel AI SDK (AI SDK 7 canary).
• Harness pages: Claude Code Harness, Codex Harness (OpenAI) — adapters wrap these runtimes; Cursor Harness not in initial adapter set ("more coming soon").
• Sandboxes: ties to Agent Sandboxes framing; Vercel Sandbox via @ai-sdk/sandbox-vercel (see also Browser Testing Skills skill for microVM browser automation).
• Kevin projects on AI SDK (Loop, Lumachor) can embed harness-backed coding agents without bespoke integration per harness vendor.
• Community analog over the open Agent Client Protocol: Spawn Agent (Aiden Bai) exposes 8 local coding agents as AI SDK providers — same "embed any agent" goal, different lineage (ACP vs first-party adapters).

Status

Experimental. Vercel explicitly welcomes feedback while the API stabilizes. Do not treat adapter package names or config shapes as frozen until GA. Source: Vercel changelog, 2026-06-12

---

Timeline

• 2026-07-04 | Source-reviewed the Vercel Developers Deep Agents/OpenCode bookmark and recorded current harness package versions: @ai-sdk/harness@1.0.18, @ai-sdk/harness-deepagents@1.0.17, and @ai-sdk/harness-opencode@1.0.18. Source: X/@vercel_dev, 2026-06-26; Source: npm registry, 2026-07-04
• 2026-07-04 | Source-reviewed the original Vercel Developers HarnessAgent row and refreshed the initial adapter snapshot: @ai-sdk/harness@1.0.18, @ai-sdk/harness-claude-code@1.0.18, @ai-sdk/harness-codex@1.0.19, @ai-sdk/harness-pi@1.0.18, and @ai-sdk/sandbox-vercel@1.0.18. Source: X/@vercel_dev, 2026-06-12; Source: npm registry, 2026-07-04
• 2026-06-26 | Added Deep Agents and OpenCode adapters; updated supported harness set to Claude Code, Codex, Deep Agents, OpenCode, and Pi. Source: Vercel changelog, 2026-06-25
• 2026-06-18 | Cross-linked Agent Harness and Agent Runtime Architecture so HarnessAgent sits in the new harness/runtime ontology. Source: whole-wiki graph audit, 2026-06-18
• 2026-06-13 | Cited in Agent Machines as validation of its dual-routing "OpenRouter for agents" bet — "swap the harness like you swap models" is the same runtime-pluralism primitive, shipped by Vercel. Source: User, 2026-06-13
• 2026-06-12 | Vercel shipped HarnessAgent on AI SDK 7 canary. Initial adapters: Claude Code, Codex, Pi. Unified session + sandbox + AI SDK-compatible generate/stream; swap harness like swapping model. Source: Vercel changelog, https://vercel.com/changelog/program-agent-harnesses-with-ai-sdk, 2026-06-12; AI SDK harness docs, https://ai-sdk.dev/v7/docs/ai-sdk-harnesses/overview, 2026-06-12
• 2026-06-12 | Guillermo Rauch (@rauchg) announced the launch: "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." 508 likes, 431 bookmarks. Source: X/@rauchg, 2026-06-12
• 2026-06-12 | @vercel_dev posted the API shape — new HarnessAgent({ harness: claudeCode, sandbox: … }), sandboxed sessions + AI SDK-compatible streams. 635 likes, 546 bookmarks. Source: X/@vercel_dev, 2026-06-12