The fix for AI code review isn't more reviewers. It's making the code worth reviewing before it ever reaches one.

Addy Osmani (Director of Cloud AI at Google) just wrote a sharp piece on this, and the numbers in it are brutal. Review duration up 441%. Zero-review merges up 31%. Defect rates climbing from 9% to 54% (Faros AI). Roughly 4x the raw code output for about 12% more delivered value (GitClear). The gap between those last two numbers is review work. We poured machine-speed output into a system built for human-speed judgement, and the whole system is now choking downstream. https://addyo.substack.com/p/agentic-code-review

Almost every proposed answer points at that downstream. Run two AI reviewers instead of one. Add mutation testing. Tighten the gates. Sample more PRs. All sensible. All necessary.

But notice what they have in common. They're all detection. They all assume the work arrives as a black box and the job is to catch what's wrong after the fact. So as agent output grows, you just keep adding reviewers to a pipe that's already overflowing. Detection scales linearly with volume. It never gets ahead of it.

The harder, more durable move is upstream. Osmani names it himself: the reviewer is often the first human to ever read the code, reconstructing intent that was never captured. That's the real cost. So capture it. Make the change arrive review-ready by default — the intent it was built for, the boundaries it stayed inside, and the evidence of what was already verified, attached to the work itself.

A reviewer who can see why a change exists and what was checked isn't slower at machine speed. They're faster.

We made writing code cheap. The next win isn't catching more bad output. It's producing work that's trustworthy on arrival.

Detection scales with the volume. Prevention changes the slope.