Reviewing AI-generated code

How to review a change that came from a coding agent rather than a colleague: why fluent, plausible-looking code needs a different review posture than human code, what to check before ever opening the diff, how to catch scope drift, missed edge cases, architecture mismatches, security and privacy gaps, bad dependencies, hollow tests, and invented APIs, how to use AI itself as a second reviewer without trusting it circularly, what to run before approving, what makes a change maintainable past the moment it merges, and when the right review outcome is a rewrite instead of another correction round.