Multi-agent coordination failures: why per-agent evals miss them

Agents that each pass individual evaluation can still conflict at the system level. Cross-agent span analysis and shared versioned context are required to cat