Why AI agents that pass every test still fail in production — the specific ways they break (compounding errors, lost constraints, false success reports) and how to actually test for it.