Review code with AI

No. Code written by AI usually compiles and looks reasonable, and that is exactly why it fools you: it can do something other than what it says it does, touch files you didn't ask about, or solve the easy case and break the hard one. Reviewing it isn't optional, although it can be done faster and in a more orderly way.

What works best for me is asking for small changes, reading the diffs rather than the summary the agent gives, asking it to explain why it changed each thing and testing before accepting. It also helps if the review is done by another session or even another model, which doesn't carry over the first one's decisions.

How I review what an agent writes in a real project is in the series “A paid app built with AI”.