Utility Forge is an experiment in autonomous software delivery: can three AI agents with fixed roles ship software to production on their own, every day, with no human in the loop? It turns out they can.
The three agents
- Ava PO (Product Owner) — generates and scores tool ideas each morning, then opens a GitHub issue with acceptance criteria for the winner.
- Eve SE (Software Engineer) — picks up the issue, implements the tool, runs the tests, and opens a pull request.
- Nora QA (QA Reviewer) — validates the PR against the original acceptance criteria and auto-merges when everything passes, triggering the deploy.
GitHub as the runtime
There is no shared runtime, message queue, or orchestration framework. The
agents coordinate entirely through GitHub primitives: issues, labels, comments,
and repository_dispatch events. All state lives in the repository, so any
issue doubles as a full audit trail of what happened and why.
An hourly watchdog re-dispatches any agent that has been idle too long, which keeps the pipeline from silently stalling after transient failures.
What it has shipped
Small, focused, no-install developer tools — a JSON formatter/minifier, a Markdown table builder from CSV, a SQL pretty printer, a cron expression explainer, and more, growing by one every day. The scoring formula Ava uses deliberately favors exactly this shape of tool: high value, low effort, and a novelty penalty so she doesn’t repeat herself.
The biggest lesson from running it: the Product Owner prompt matters more than anything downstream, because Ava’s acceptance criteria become Nora’s QA checklist. The full write-up covers the coordination design and the failure modes.