Five Parallel AI Builders, One Contract, 25 Review Fixes
On September 26 I had five builder agents work from one markdown contract that said which files were theirs. When P0, the first local version, was done, it ran at 127.0.0.1 with 224 passing tests and no requests leaving the machine. Before that, a review pass ended in 25 fixes, and I think that number says more about parallel agent work than the test count does. This is part of a series about…
On September 26, five builder agents collaborated on a single markdown contract outlining file ownership for the project. The first local version (P0) ran successfully on 127.0.0.1, passing 224 tests with no outgoing requests. A subsequent review identified 25 fixes, highlighting the challenges of parallel agent work more than the test count itself.
The series discusses the development of Refmade, a desktop app that organizes Claude Code and Codex sessions as team-based projects, with each agent taking on specific roles. The process involved breaking down tasks, freezing existing tests, and establishing strict rules to ensure consistent work across the builders. A key requirement was that builders modify only their designated files by editing the contract first, preventing unintended changes to others' work.
The app runs on a loopback port accessible only to the machine, with a unique key generated for each session to protect the interface from unauthorized access. The review process involved multiple angles and refutations, resulting in 25 fixes before final implementation.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
