Multi Agent Coordination Bench
Bespoke Labs has released Multi Agent Coordination Bench, a new benchmark comprising twelve natively multi-agent long-horizon tasks designed to evaluate the coordination capabilities of frontier AI models. Evaluating models from OpenAI, Anthropic, and Google, researchers found that current frontier agents coordinate poorly in teams, with performance gains primarily stemming from additional agents rather than effective communication. The findings indicate significant headroom for improvement in score, compute efficiency, and multi-agent coordination. Open article →