About
agent-harness-kit is a CLI tool that solves one of the core problems with AI coding agents: they’re stateless, uncoordinated, and have no memory across sessions.
Running npx @cardor/agent-harness-kit init in any repo generates the full infrastructure for a multi-agent system — SQLite database, typed config, per-agent instruction files, a built-in MCP server, and a health check script — in under a minute.
The harness ships with four specialized base agents, each with explicit permission boundaries:
- Lead — orchestrates everything. Picks tasks from the queue and coordinates the other agents.
- Explorer — read-only. Maps only the code relevant to the current task before anyone touches the repo.
- Builder — writes to
src/andtests/only. Implements what the Lead assigns. - Reviewer — gatekeeper. No task moves to done without passing tests and Reviewer approval.
Problems it solves:
- Session amnesia — agents share persistent state via SQLite. The Lead always knows what was done and what comes next.
- Agent collision — roles are enforced with file-level permission boundaries, not just suggestions.
- No verification — the lifecycle gates "done" status behind Reviewer approval and green tests.
- Context bloat — Explorer maps only relevant code per task, not the entire codebase.
- No specialization — each agent has a single responsibility instead of doing everything solo.
Works with Claude Code and OpenCode. Uses node:sqlite (Node 22+) or bun:sqlite — no external database or Docker required. Available on npm as @cardor/agent-harness-kit.
Tech used
- TypeScript, Node.js
- SQLite (node:sqlite / bun:sqlite)
- MCP (Model Context Protocol)
- CLI (interactive prompts via clack)
- Claude Code, OpenCode