For the past few months I've been using Cursor or Claude Code on most of my PRs. They genuinely speed things up: endpoints, repositories, test scaffolds often land in ten minutes. But as diffs grow, I can't review at the same pace — sometimes it feels slower than before.
Agents shine on repetitive patterns: migration drafts, MediatR refactors, README or ADR drafts. They work great as draft generators. The pain starts at domain invariants — money, inventory, idempotency. Agents write plausible code; production usually breaks where invariants live.
The most common review trap is trusting the agent's test scaffold. A happy-path test like CreateOrder_ReturnsOk helps but isn't enough. If the diff lacks negative cases — insufficient stock, duplicate idempotency keys, concurrent orders — I don't merge. Reading the agent's PR summary isn't review; reading the diff line by line is.
My pre-merge list stays short: if tests are missing, is there a reason; are logs, metrics, and traces in place; are auth and transaction boundaries clear? Agents accelerate; ownership stays with the engineer.