Taming the eager intern: a workflow harness for AI coding agents

What it does

AI DevKit is a TypeScript CLI that scaffolds a repeatable engineering workflow across AI coding agents. One .ai-devkit.json generates per-agent configs, slash commands, and MCP wiring so Claude Code, Cursor, Codex CLI, Gemini CLI, and others all follow the same phases: requirements, design, planning, implementation, tests, and review. It also provides a local SQLite memory store and a TUI dashboard for monitoring running sessions.

The interesting bit

The project treats agent behavior as infrastructure rather than prompt engineering. Instead of static rules files the agent might ignore, it installs operational gates: /new-requirement blocks coding until the problem is clarified, verify demands fresh test output before marking tasks complete, and /code-review audits diffs against design docs. The “agent console” even lets you pipe session output through Telegram to check on long-running tasks from your phone.

Key highlights

• One config reconciles commands, skills, and MCP setup across 10+ agents (Claude Code, Cursor, Codex CLI, Gemini CLI, Copilot, and others)
• Eight built-in skills address specific failure modes: dev-lifecycle, verify, memory, tdd, structured-debug, document-code, simplify-implementation, technical-writer
• Local SQLite memory persists decisions and conventions across sessions; no hosted service or telemetry
• ai-devkit skill add pulls additional skills from 30+ publishers including Anthropic, Vercel, Supabase, Microsoft, Google
• Remote control surface: send prompts to running sessions, pipe output, or route through external channels

Caveats

• Remote control only works with Claude Code, Gemini CLI, and Codex CLI; Cursor, Copilot, and most others are setup-only
• Windsurf, Kilo Code, and Roo Code support is marked “testing”
• The README notes this “raises the floor on process, not on raw capability” — bad models stay bad

Verdict

Worth a look if you’re running multiple AI agents across projects and tired of re-rigging CLAUDE.md or .cursor/rules by hand. Skip it if you only use one agent with a simple workflow, or if you expect the tool to compensate for a weak underlying model.