Updated October 12, 2026: this started as one long post. It’s now this overview plus four shorter parts, with what’s changed since: support email that runs without me at the laptop, batched manual testing, and the agent that audits the process itself.
This is an experiment, and I picked a safe place to run it. RepBunny keeps every workout on the phone it was logged on. There are no user accounts, no shared database, and no servers of mine holding anyone’s data. So when I let AI agents merge code with no human code review, the worst a bad change can do is break the app on a phone, and a person still tests every release on a real phone before it goes public. It can’t leak anyone’s information, because I never have it. I wouldn’t run this setup on an app that held other people’s data.
The short version
RepBunny is a strength training log for iPhone and Android. I’m the only developer. Behind it is a set of AI agents (Claude Code, mostly) that each have one job: turn my rough notes into specs, build the next most valuable issue, test it, ship it, review the app every month, and ask me when something needs a human call.
None of that is one giant prompt. It’s about ten small skills, a dozen GitHub Actions workflows, and a label scheme on GitHub Issues that every agent reads and writes. The issues are the team’s shared memory. If it isn’t on an issue, it didn’t happen.
- 01MeJot an ideaA one-line issue from my phone, or a bug from a tester.
- 02AIRefineAn agent checks it against the code and rewrites it as a spec with effort and value scores.
- 03AutoDispatchA plain script sees work in a lane and starts that lane's agent.
- 04AIBuildThe agent claims the top issue by ROI and writes the change.
- 05AutoGateFull build plus 677 end-to-end tests. Red means fix or stop.
- 06AIMergeGreen means the agent merges its own PR and closes the issue.
- 07AutoTell meA "shipped" email lands with what changed and how it was tested.
- 08MeReleaseI cut a release candidate and smoke test it on a real phone before it goes public.
The series
Each part stands on its own and takes four or five minutes.
Part 1: From a one-line idea to the right AI agent. How one-liners typed at the gym become scored specs, how four monthly review agents fill the backlog, and how a dispatcher with no AI in it sends each issue to the cheapest model that can handle it.
Part 2: No human code review: why I trust tests instead. Agents merge their own pull requests after a full end-to-end gate. A person still tests every release on a real phone, and the checks only a real phone can do now come in batches.
Part 3: My AI agents ask me questions. I answer from my phone. The ask-the-owner loop, support email that triages itself, and a sample conversation from the first day it ran on its own.
Part 4: My AI team audits its own process. A monthly DevOps reviewer that reads the bill. It found an Actions budget on pace to run out, and docs that cost 43,000 tokens per run.
If you want to try this
A few things I’d do again on day one:
- Pick one source of truth and make every agent read and write it. For me that’s GitHub Issues and labels. Agents forget everything between runs. The issue tracker doesn’t.
- Invest in the test suite before the automation. The gate is what lets you skip code review. Without it, you’re just merging faster in the dark. And keep a human between the gate and your users.
- Keep the glue dumb. The dispatcher, emails, and ask-the-owner flow have no AI in them. They’re cheap, predictable, and easy to debug.
- Write the rules down in the repo. Every decision I make ends up in a docs file the agents read at the start of each run. That’s how a choice I made on Tuesday still holds on Friday.
- Plan for yourself as the bottleneck. The more the agents do, the more your own to-do list becomes the thing everything waits on. Make your part fast and phone-sized, and batch the parts that can’t be.
RepBunny is in a small private pilot and coming to iPhone and Android. You can read about it at repbunny.app. Questions about any of this? Email me at me@deanphillips.dev.