Long-running agents need checkpoints because a person’s goals can change, new evidence can appear, and systems can drift.