What they did
Rodrigo Lemela Duarte runs a small software team where most engineers are agents: several Claude Code engineer agents, each on its own clone and branch, one Claude Code manager agent, and Codex reviewers. They coordinate in a shared chat room with the people on the team, a product the team itself builds. Only the manager agent merges. It merges to staging by script, one batch at a time, never while a pipeline runs. Before each batch it runs the same verify script as CI, checks every pair of PRs for conflicts with git merge-tree, then verifies the combined tree. It also sets the landing order, and migration numbers follow that order. Production, cost and credentials stay with a person.
How it went
In the week described, an engineer agent reported green because npm test passed, but the manager agent ran CI's verify script and found 17 type errors from a type declared twice. Two PRs collided in the deploy doc, and one would have carried a new database migration into the next day's release, so it was split. A migration numbered 167 was renumbered because staging was at 168 and open PRs held 169 to 171. One batch still went red after merge on a 100 ms date-field focus bug the gate could not see. A Codex reviewer caught that a git hunk header in a .env diff could expose a secret.
Worth knowing
One Codex reviewer for all engineer agents became a bottleneck; the team moved to three reviewers with fixed engineer pairs, and the second review round stays with the first reviewer. The post is written by the makers of the chat tool the agents use, and gives no figures on how many PRs went through.