I Let a Claude Agent Review My Code. It Was Harsher Than My Lead.
The agent flagged three things my human reviewer missed, and none of them were style. It changed how I pre-submit.
Before sending a pull request, I ran it past a Claude agent configured to review like a skeptical senior. It came back with three issues my human reviewer later agreed with: a race condition under load, a missing null check on an external response, and a function that duplicated logic two files over. None were formatting. All were the kind of thing that ships and bites later.
What surprised me was the consistency. A human reviewer has a mood and a deadline; the agent reads the same way at 2am. It doesn't skip the boring file because it's tired. That reliability is the actual value, more than any single catch.
It's not a replacement. The agent missed the why behind a weird workaround and suggested removing it, which would have broken a legacy integration. I caught that because I know the history. So the workflow is agent first, human last, and the human step stays non-negotiable.
I now run the agent on every change before the human sees it. It filters the obvious and the structural, so my reviewer spends their attention on the judgment calls. The review got cheaper and better, which is rare to say about a process change.