Stanford benchmark finds top AI agents score only 82.2/100 as teammates
StanfordUniversity · hf · 2026-11-03
Stanford with MIT and Oxford introduce the Collaborative Intelligence Benchmark: 240 tasks across 8 teamwork dimensions, from refusing unsafe shortcuts to checking a weaker model's work, drawn from aviation accident reports, clinical cases, coding tasks, and negotiations with programmatically verifiable outcomes. Models are scored only on tasks they can solve alone to isolate teamwork from raw capability. Across 18 closed- and open-weight models, every one shows clear teamwork weaknesses—the best scores just 82.2/100. Frontier models almost always refuse unsafe shortcuts yet cave under interpersonal pressure and fail to reach the best overall deal.
More from coding & agent
- LetterBox adds granular API key permissions, agents can create Drafts over MCP — rchardkovacs · 2026-11-03
- Tsinghua study: continual harness self-evolution causes severe catastrophic forgetting in coding agents — Shangziqi Zhao · 2026-11-03
- Vantage: open-source Windows app lets AI agents drive your real desktop, 80-step task in ~8 min — AKGAMING1234 · 2026-11-03
- Reward hacking risks when agents optimize inference, new article warns — AccBalanced · 2026-11-03
- Build a Media Agent workshop: CPU-based agent calling GPU tools on demand — Al_Grigor · 2026-11-03
- Factory rebuilt its Droid coding agent to run inside customers' own clouds, plans to open source the compute layer — AI Engineer · 2026-11-03