Most AI coding tools give you a chat window and a single agent with no memory. That breaks down fast when you want: - Multiple specialized agents working in sequence - A record of what was tried across sessions - A health check that must pass before agents touch code - The ability to switch AI providers without losing context<p>agent-harness-kit is a thin layer you add to any project. It runs a local MCP server (stdio), backed by a SQLite file. Agents pick tasks from a backlog, claim them atomically, log every action and file they touch, and can only close a task after a health script exits 0.<p>The 4-agent workflow (Lead → Explorer → Builder → Reviewer) is opinionated by default but fully customizable — the agent definition files are plain Markdown and are never overwritten after init.<p>No cloud. No native compilation. Zero deps beyond what Node 22 ships with.<p>npx ahk init<p>Source: <a href="https://github.com/enmanuelmag/agent-harness-kit" rel="nofollow">https://github.com/enmanuelmag/agent-harness-kit</a>

[flagged]