Collaborative Intelligence Benchmark Tests AI Teamwork

This is a valuable shift from asking whether agents can solve tasks to asking whether they make teams better. Multi-agent builders should treat collaboration as its own capability, with dedicated evaluation and failure analysis.

• –Tests safety pressure, information sharing, conflict, initiative, and compromised handoffs
• –Runs solo, normal-team, and disrupted-team versions to isolate teamwork failures
• –Shows strong models can still inherit mistakes or accept inferior outcomes from pushy teammates
• –Public tasks and code make it practical for teams to add collaboration checks to model evaluations
• –The scripted-teammate setup improves reproducibility but may not capture messy real-world teamwork