🤝 Collaboration Benchmarks

Real multi-agent execution — separate LLM calls per agent, multi-round interaction, information decay measurement