Every agent framework has the same safety net: when in doubt, ask the human. The data says that net has holes. Developers approve over 90% of agent permission prompts, and in one controlled study reviewers caught fewer than one in seven dangerous commands. None of this is new. In 1948 the RAF learned that radar operators started missing targets after half an hour on watch. In 1972 a flight crew flew a working airliner into the Everglades while all three of them troubleshot a burned-out bulb. In 1983 Lisanne Bainbridge explained why: automate the routine work, and the human is left with the one job humans do worst, which is monitoring. This talk traces that line from radar rooms and cockpits to MFA fatigue and agent approval prompts. It then lays out what to build instead: containment that removes routine prompts, approvals that ask about intent rather than commands, and telemetry that tells you when your reviewers have stopped reading.