Case study · 06
Ralphy
An autonomous AI coding loop: a CLI that runs coding agents again and again until the task, or the entire PRD, is complete.
What we built
Ralphy is an autonomous coding loop: you give it a task or a full PRD, and it
runs an AI coding agent against it, again and again, until the work is
actually done. It ships as an npm CLI (ralphy-cli) and an equivalent bash
script, and supports six agents: Claude Code, Codex, OpenCode, Cursor Agent,
Qwen, and Droid.
Two modes cover most real work:
- Single task:
ralphy "fix the auth bug"and walk away. - PRD mode: point it at a task list (
ralphy --prd tasks.md) and it works through the document item by item.
How it's built
The hard part of autonomous coding isn't starting the agent; it's keeping it
on the rails across many unattended iterations. Ralphy handles that with a
per-project configuration (.ralphy/config.yaml) that travels with the repo:
- Rules
- Standards the agent must follow on every task, like "use TypeScript strict mode" and "use server actions, not API routes", so they hold without a human restating them each run.
- Boundaries
- Glob patterns the agent must never touch, protecting legacy code and lockfiles from well-meaning refactors.
- Project commands
- Test, lint, and build commands the loop runs to verify work before calling a task complete.
- Agent-agnostic
- One harness, six interchangeable agents, so teams can switch models without switching workflows.
Why it matters
With 2,900+ GitHub stars and an active community, Ralphy is one of the most-adopted autonomous agent harnesses in public. It's also a working answer to the question every client asks us: how do you let AI do real work without losing control? Rules, boundaries, and verification, encoded in the tool itself.