Apate’s persistent scam decoys, OpenAI’s self-damaging research model and Nadella’s proposed emergency brake expose the double edge of agents that keep trying. Builders should make stopping conditions and human interruption part of the product’s definition of success.