Agents.Code — Autonomous Agent Coding

Autonomous coding systems built on the Claude Agent SDK. The project is split into two complementary subsystems with deliberately different design goals, over a small shared core.

• More on the how and why: blog.

Subsystems

Subsystem	Shape	What it does	Docs
src/harness	finite pipeline	spec → steps → plan / generate / evaluate → done. Compiles an editable feature spec into a working app, one verified step at a time.	src/harness/README.md
src/loop	perpetual cycle	discover → distribute → verify → record → decide. A discovery-driven operating discipline kept separate from the spec→app compiler.	src/loop/README.md

src/shared holds the common core both reuse — logging, auth, and SDK-stream consumption — and is otherwise dependency-free between the two.

src/
├── harness/   spec → app compiler          → src/harness/README.md
├── loop/       perpetual discovery cycle    → src/loop/README.md
└── shared/     logging · auth · SDK stream

Quick Start

# Install dependencies
npm install

# Install Playwright (used by the harness's generator + evaluator)
npx playwright install chromium

# Run the harness: scaffold a draft spec, then build from it
npx tsx src/harness/index.ts "Build a task management app with kanban boards" --output-dir ./kanban
npx tsx src/harness/index.ts ./kanban/artifacts/<auto-slug>/spec.md

See src/harness/README.md for the full CLI, options, artifact layout, and how the per-step loop works.

Authentication

• Signed into Claude Code locally — nothing to configure. The SDK subprocess inherits your existing session via settings.json.
• Explicit auth — set one of the auth variables in .env (API key, OAuth token, Bedrock / Vertex / Foundry, or an apiKeyHelper). See .env.example for the full list.

References

• Harness Design for Long-Running Apps — Anthropic Engineering
• Claude Agent SDK — Official Docs