Agentic Doctor is a recursive codebase doctor that pairs an advisor model with an executor model. The advisor chooses one bounded next fix and emits a /goal; the executor edits the target repo, verifies the change, and hands the result back for review.
The hackathon objective was to build something with two AI roles, an advisor and an executor, working through Vercel AI Gateway. Agentic Doctor applies that to a recursive coding loop: the advisor keeps the next fix small and testable, while the executor implements it, verifies it, and hands the result back for review. It was inspired by react-doctor and how it can be used as an eval inside a Ralph loop for improving a React codebase.
Agentic Doctor was built during America's Next Top Model in San Francisco.
- Vercel AI Gateway routes advisor and executor calls through configurable models.
- Advisor loop inspects the target repo context and emits either one bounded
/goalor/done. - Executor loop receives only the current goal, edits the target repo, runs verification, and reports changed files, results, and risks.
- Recursive review sends the executor result back to the advisor so it can choose the next goal or stop.
- Direct Gateway mode runs both roles through the built-in model client.
- Harness mode runs Claude Code as the advisor and Codex CLI as the executor with local coding tools.
- TypeScript CLI provides model selection, target snapshotting, logging, and loop controls such as
--max-roundsand--allow-done.