If your team uses coding agents, you've probably noticed that code review got slower. The queue fills faster, the PRs are harder to read, and you're spending more time on review without getting more confident in what you're merging.
AI-generated code is idiomatic, well-named, consistent with the surrounding codebase. When there's a problem, it's in the logic — you have to reconstruct what the code was supposed to solve, not just read what it says. That's slow work. Pushing back on the model doesn't help: it agrees with everything, which tells you nothing about whether you're right.
The cost is in where feedback lands. A round of feedback on a PR — reloading context, rewriting code already written, waiting for CI — cost us 16 hours on average. On a plan, before any code exists, the same feedback costs under an hour.
So we moved to reviewing plans. A developer writes what they intend to build, a teammate challenges it, the agent builds against the agreed scope. We added a check at the end that compares the diff against the approved plan, because agents drift and you only catch it if you're looking.
Thirteen of our first 165 plans were dropped before any code was written. That number doesn't show up in any metric because nothing happened — the intent was wrong early, caught early, stopped. The work that did go forward moved through review 4.7× faster.
The code review stage is the wrong place to find out if you're building the right thing. The plan is cheaper, and earlier.
The data behind this — 22,000 developers, and why engineering maturity doesn't protect you — is in AI is breaking engineering teams.