Inigo AnguloContact
Work01How I think02Lab03Writing04
Writing

445 PRs in seven weeks: what a one person team with AI agents actually looks like

Inigo AnguloNovember 4, 20267 min readAI agent teams

Between September 16 and November 2, 2026, I built and shipped a production platform for car collections with no engineers on payroll. The GitHub API says 484 pull requests were opened, 445 merged, 38 closed without merging. 1,456 CI runs, 1,026 green. Fifteen numbered releases to production, each with a security report and one human approval. I want to describe what that looked like from the inside, because "I built it with AI" has started to mean anything, and the useful version is specific.

17 daysfirst commit to first production deploy
445pull requests merged
1,456CI runs, 147 red
15gated releases

The team, and why it is a team

The work was done by eight AI agents with fixed roles: a lead that decomposes tickets and integrates, workers for frontend, backend and data, a reviewer, a security agent that runs a full penetration test before every promotion, a design agent with the only write access to Figma, and a verifier that only reads. Each one has a mandate written in a file, with what it may touch and what it must never touch. None of them has production credentials. That last sentence is the whole architecture.

Calling it a team is not marketing. The roles matter because the failure modes are the same as with people: a worker that starts from the wrong branch, a reviewer that approves its own work, a check that looks green because it did not run. Every rule in the repository exists because one of those happened and cost a day.

What the human did all day

Not code. I wrote tickets with acceptance criteria, answered decisions and approved releases. Every open question the agents hit became a numbered decision with lettered options and a recommendation; I answered with an id and a letter, "014-3 B", and the answer was written into the repository with the reasoning. Over seven weeks that produced more than 140 recorded decisions. That log is the product's memory, and it is why a new agent session can pick up where the last one stopped.

The other thing I did was say no. 38 pull requests were closed without merging. 147 CI runs went red. A red check on the exact commit blocks the merge, whoever asks, and the "whoever" includes me on a Friday evening wanting the feature out.

The rhythm

Week one was infrastructure: database, auth, row level security with tests that run as the role that must not pass, the CI pipeline. Day 17 was the first production release. From week four the rhythm settled at roughly forty tickets a week. A typical ticket, from a real day: written at 14:35, pull request opened by the lead at 16:56, fourteen checks green, pen test PASS at 17:20, merged, promoted at 19:50, served from the production domain at 19:54. Five hours, with a human at two of the gates.

What was not fast

Judgment. Deciding what the product should be took as long as it always has. The data model was argued over for days. The agents are very fast at building the wrong thing, which is why the first week was spent on invariants: which rules live in the database, which in the application, and the test that proves each one.

And verification. Reading a CI run instead of trusting its colour. Opening the deployment by its URL and looking for the content instead of asking the provider whether it was "READY". The gates where a human looks are the product; the agents are what make them affordable.

What I would tell a team adopting this

This post is part of a series: the hub is AI agent teams for software development, and the gates are in Human in the loop for coding agents. The full case, with the evidence from the GitHub API, the data model, the promotion flow and the day the security gate said no: Acervo, the case study. If you want this for your company: AI build and deployment.

Working on something like this?

This is the method behind how I build: AI native development with human gates. Discuss a product idea → Or email me.