AgentWorld benchmark: best multi-agent team solves just 52% of tasks
OpenAgents, Columbia University, the University of Pennsylvania, Seoul National University and Penn State built AgentWorld, a benchmark for multi-agent collaboration that puts LLM-driven agent teams into an MMORPG sandbox with 100 hand-designed tasks plus 100 enhanced variants covering division of labour, communication and resource handoffs. Under a shared prompt template, tool set and interaction protocol, Gemini 3 Flash, Claude Haiku 4.5, GPT-5 Mini and DeepSeek R1-70B scored 52.0%, 45.0%, 36.0% and 20.0% on the main task set, with the enhanced set topping out at 26.0%. The paper also introduces a Causal Collaboration Effectiveness (CCE) metric that traces which actions and messages actually contributed to the shared goal.
It turns “can agents collaborate?” into a measurable capability, suggesting the bottleneck for multi-agent systems may lie in information sync and handoffs rather than single-agent skill.
个体能力强大的AI Agent正不断涌现。从编写代码、开发软件，到检索文献、辅助科学研究，它们能独立完成的工作越来越多。
但随着这些「超级个体」变得更强，一个同样重要的问题也浮现出来：它们能否合作？
回看人类社会，许多复杂成就都依赖一群人的分工与协作。假如我们要组建一家公司，制造火箭、登上月球，首先要考虑的，恐怕不是寻找一个无所不能的天才，而是如何组织一支团队：需要哪些专业人才，怎样分工，如何共享信息，又如何让不同环节的工作衔接起来。
从科学发现到大型工程，个体能力固然重要，组织和协作也决定了我们能够走多远。
这让我们开始思考：当大模型和Agent的个体能力持续增强时，它们的协作能力发展到了什么程度？面对一个共同目标，它们会不会主动交换信息、协调分工？当多个智能体一起工作时，能否真正发挥各自的优势，让复杂任务更容易完成？
#OpenAgents#Columbia University#University of Pennsylvania#Seoul National University#Penn State#AgentWorld#Gemini 3 Flash#Claude Haiku 4.5#GPT-5 Mini#DeepSeek R1-70B