Eval suites that only grade happy paths lie. I add adversarial prompts and empty tool returns so the agent fails in staging not on a customer ticket.