The first weeks of working with coding agents feel like speed. The second weeks feel like reading. Somewhere in between, teams discover that the bottleneck moved: generating code is cheap, and deciding which of it deserves to ship is the craft.
Read the diff you didn't write
Machine-made diffs read differently. They are locally plausible and globally ambitious: every hunk makes sense, and taken together they quietly widen scope. The review move that catches most of it is unfashionably simple — name the job in one sentence, then strike everything the sentence does not ask for.
- Diff the diff against the brief, not against your memory of it
- Run the acceptance checks before forming an opinion
- Treat unrequested beauty as scope creep
- Approve fast or reject clearly — half-decisions re-enter the loop
Refactoring made existing code better. Review decides what deserves to exist at all.
From the peer-review loop
None of this is a downside of agent work. It is the compounding part: every sharply rejected diff teaches the loop what good looks like.