Ninety days with AI agents, measured like a program.

Between late July and late October 2026, I built Virabelo, a ten-repository drone operations platform, working alone with AI coding agents. I measured it the way I would measure a program I was accountable for: from the git history, the pull requests, the CI runs and the bug tracker, not from how fast it felt.

What the agents produced

• 98% of the 1,562 commits in those 90 days were authored or co-authored by AI agents.
• About half of all new lines were tests.
• 768 pull requests merged across the ten repositories.
• Across the project, the median logged bug had a code change on the main branch within three hours.

That is the part everyone talks about, and it is real. One person with agents produced the volume of a team. The more useful findings were in the controls around that volume.

The gates were leakier than I thought

It is easy to assume every change goes through CI before it merges. The record says otherwise, at least early on. In August, most merges ran no CI at all. By October, nearly every merge ran CI and most waited for it to finish.

All merged pull requests across ten repositories. *Share of merges with CI where every check finished before the merge. The last three days of July (22 merges) are left out.

Agents move fast enough that a gate you believe in but don't enforce gets skipped dozens of times a week. Adding CI to every repository closed most of the gap. The rest closes only when CI is a required check, so nothing merges until it passes.

Review left no record

None of the 768 merged pull requests has a recorded review. Review happened inside the working session, but nothing in the record shows it. To an auditor, a prime contractor or your own quality lead, review that leaves no trace did not happen.

A solo builder can't approve their own pull request, so the answer is a reviewer agent, separate from the one that wrote the code, that posts its findings on every pull request, followed by a short human sign-off before merge.

The tracker drifted

Agents fixed bugs quickly and rarely closed them. Worse, when a new defect turned up nearby, it often got added to the old ticket instead of a new one. At the end of October, 127 of 369 open bug tickets were already named in merged code, and 83 of those had picked up comments after the code landed.

The rule I now run by: one defect, one ticket, and closing the ticket is part of merging the fix.

If you run engineering programs

1. Measure the cost of an accepted result, not lines or commits. Volume is the easy number and the least useful one.
2. A gate only counts if the record shows it. Required CI, and review that leaves a trace.
3. Expect the tracker to drift. Agents amplify whatever process you have, including its gaps.
4. Watch the spend. AI tools are moving to usage billing, and that spend needs to land on the program that caused it, not in overhead. That is what I am measuring next on this same program.