The timeout mismatch example is a good one because it's not even a case of the AI being wrong about syntax, it's wrong about a fact it never had access to. The one I keep running into in agent-generated integration code is subtler: retry logic that assumes an HTTP-level failure contract when the actual downstream API returns 200 with the real error buried in the response body. Nothing in review catches it because the retry policy looks completely sane in isolation, and it only shows up once you're staring at logs full of successful-looking calls that quietly did nothing. Writing down the tribal facts before merging is a good habit, but I'd add a cheap mechanical check on top: grep the generated code for every place it treats HTTP status as ground truth and cross-check that assumption against a real captured response from that specific downstream, not the API docs.
Tarek Mostafa
sharing great ideas on developers experience