Agent asks before it deletes prod

Your migration agent runs in its own container, working through 200 schema changes. Before any DROP TABLE or DELETE FROM that touches more than 1000 rows, it pauses and POSTs to hoody-notifications: urgency: critical, expire_time: 0, body: "DROP TABLE legacy_orders: confirm?".

The notification surfaces on the display tab you keep open in your phone browser. You tap, you read the SQL, you reply yes or no through the agent's HTTP endpoint. The agent resumes. If you don't respond in 30 minutes, it skips the destructive step and moves on.

Most agent failure modes aren't "the LLM was confused." They're "the agent did the irreversible thing without checking." A POST /notify and a confirm endpoint costs 50 lines of code and gives you a circuit breaker that runs at human speed when it matters.

The agent doesn't have to be perfect. It has to know when to ask.