What they did
The author, who writes as Sesisnando, built an open source tool called guild on top of Claude Code and Codex and uses it for daily work. One coordinator agent takes requests and tickets and splits them into tasks. Each task goes to a worker agent in its own git worktree, four or five at a time. Before any code is written, the worker posts a plan to a local web page and waits for the author to pick an option. Before a pull request opens, a reviewer agent reads the diff, fixes safe findings, and the project's tests, lint and type checks must pass. Work is routed by model: Sonnet builds most features, Haiku takes small fixes, Opus plans and reviews, Fable handles the hardest questions, Codex does images and long mechanical work, and OpenRouter models are the fallback when a provider hits a usage limit. It only comments on GitHub or Jira when tagged.
How it went
The author says it is now how they ship code every day, and that the main change was moving from typing to deciding. Their cost page showed Fable finishing 15 tasks for almost the same spend as Sonnet on 135, so they now keep Fable for hard problems. Things that failed: more than five agents on one laptop got slow and hard to follow, early plans were far too long for small fixes, and some passing tests proved nothing, so the review step now checks the tests too.
Worth knowing
Five parallel agents was the practical ceiling on one laptop. The plan approval gate is what the author credits for trusting the output. Keeping a backup model per task type avoided stalls when one provider hit its limit. Costs are tracked at API prices, so watch spend per model rather than per task. The code is on GitHub (SesisnandoLRNeto/guild).