Since agents became first-class engineers on our team, commit and PR volume is up roughly 32× at flat headcount. That number is real, and it is also the least interesting thing about what happened.
The interesting thing is what it did to the job. When agents write most of the code, the scarce human activity is no longer writing — it is deciding what should exist and whether what exists is right. Review went from a courtesy you performed on a colleague’s branch twice a day to the actual substance of the role.
Three concrete changes, with the failure modes attached.
Review became the bottleneck, so we engineered it like one. We used to treat review latency as a social problem. Now it is a throughput problem: if a human can meaningfully review N changes a day and the fleet produces 30N, the pipeline is only as good as our triage. We route by blast radius — schema changes and anything touching attribution or undo get a senior human, mechanical refactors get a fast pass with automated checks. Getting this routing wrong was our first production incident of the agent era: a “mechanical” change that wasn’t.
The unit of review changed. You do not review 32× more PRs by reading 32× more diffs. You review intents, invariants, and tests. We now require every agent PR to open with the claim it is making — what behavior changed, what cannot have changed — and the review verifies the claim, not the keystrokes. When the claim is wrong, the PR is wrong, whatever the diff says.
Taste became a written artifact. A senior engineer’s judgment used to live in their comments. That does not scale to a fleet that ships while you sleep. Every recurring review comment gets promoted into a written convention the agents load before they write a line. The corpus of conventions is now one of the most valuable files in the repo — it is, literally, our engineering taste under version control.
What we have not solved: review fatigue is real, and an agent fleet is patient in a way authors never were. The honest answer is that some days the queue wins.
If reading that sentence made you want to fix it, that is roughly the interview.
— Eng lead, New York