Developer builds self-reviewing AI agent loop, shares key design lessons

A developer has released Task Hounds, an open-source local workspace where three AI agents — a Manager, Worker, and Reviewer — collaborate in a structured loop to plan, implement, and critique code. The system was built after months of experimentation and is designed to run entirely on a local machine, storing all activity in SQLite with a real-time dashboard. Key lessons included serializing tasks one at a time to prevent agents from overwriting each other's work, anchoring the original human directive to stop goal drift over long runs, and using fixed document schemas instead of shared conversation history to avoid context window bloat. The developer also found that separating the review and decision-making roles was critical, as allowing the Reviewer to assign fixes directly caused infinite negotiation loops. Task Hounds is MIT licensed and publicly available on GitHub.