How the apps get built: one developer, a team of agents and one Mac
- studio
- process
I’m building six apps, and I’m one person. That only works because most of the building is done by AI agents (Claude Code), and because of the process around them. Here is how it works and where it still doesn’t.
The loop
Work starts as a roadmap card: one small change, with a list of “Done when” boxes that can each be checked. From there:
- A long-running lead session picks the card and writes a brief.
- A worker agent takes the brief into its own copy of the repository. It writes the change, builds it, runs the tests, takes screenshots in the simulator and reads them.
- It hands back a file: what it did, the evidence for each box, and what it could not show.
- A reviewer with a fresh context reads the card, the diff and the pictures, and returns pass or changes. It is a different model from the one that wrote the code, and it never sees the worker’s explanation of why the change is right.
- After a pass and green CI, the change is merged and the card’s status is recorded.
- One line goes into a retro log: what worked, what broke, the lesson. Lessons become rules in the files the agents read.
I don’t test builds by hand. I decide what gets built, anything that can’t be undone, and anything that is a matter of taste. Visual choices come to me as two or three real screenshots.
Why the reviewer is separate
A reviewer that shares the author’s context tends to accept the author’s reasoning. Giving it only the card, the diff and the screenshots made it find things the worker’s own checks passed over:
- A regenerated data file had lost part of 216 source links, because the generator fell back silently when its cache was missing. The worker’s tests passed.
- A new layout rule cut long two-word names in every list. The sample name fit by luck, and only a picture with a long name showed it.
- A fix for one folded-phone pose removed the only Share button in another pose.
- A pixel test “proved” a fix by reading through translucent glass.
Some numbers
These cover the first four and a half days of the current setup (November 12 to 16, 2026), on two of the apps:
- 30 pull requests merged, carrying 46 roadmap cards. Related small cards are bundled.
- Review rounds per pull request: median 2, most 6.
- On 24 of the 30, the first review asked for changes.
What they don’t show: there is no baseline to compare against, no count of defects that got past review, and no cost figure. A high rate of “changes” reflects a strict protocol as well as real defects, since a “Done when” box without evidence is enough to send a card back.
What broke
The limit is the machine. One Mac mini is both the simulator host and the CI runner, and it crashed twice in one day with several simulator jobs running at once. Since then there is one lane: a single simulator job at a time, and a script that checks the machine has room before anything boots. A worker that only pushes code still counts, because its push starts CI on the same machine.
The stop rule that protects the machine was wrong three times in a row. It kept stopping healthy runs, because a simulator’s first boot alone looked like trouble. Each version was replaced using readings from the failed runs.
The lesson that cost the most: writing a rule down changes nothing. After two weeks I went back through one app’s 102 pull requests and found five mistakes that had kept happening after their rule was written, because the rule sat in a log that no worker reads. Now a lesson goes into the worker’s or the reviewer’s own instructions as a required field or a check, at every merge.
What I can’t claim yet
None of these apps has shipped, so there is no feedback from real users in any of this. Review is expensive: some small cards took four to six rounds. Some flaky tests are still unexplained. And the reviewer and the worker are models from the same company, so they may share blind spots I haven’t found.
The files
The working files are public and kept in step with the live setup: the agent definitions, the hooks, the machine scripts and the full retro log. They are in agentic-sdlc on GitHub.