New Test Shows AI Teams Fail Half the Time at Working Together

A group of university researchers just released AgentWorld, an open-source test that puts AI teams inside a video-game world — think World of Warcraft, not a spreadsheet. Each test has 100 hand-written tasks, and teams of 3 to 20 AI programs (software that can take actions on its own) must work together for 50 or more rounds to finish them. The twist: each AI acts alone and cannot see what its teammates are thinking.

The results are humbling. Four leading models took the test — Google's Gemini 3 Flash, Anthropic's Claude Haiku 4.5, OpenAI's GPT-5 Mini, and China's DeepSeek R1-70B. The best of them completed only 52% of the tasks. That's barely better than a coin flip. The failures followed clear patterns: teams stopped talking to each other, forgot the plan halfway through, or got confused about who was supposed to do what.

The researchers also invented a smarter way to grade teamwork. Instead of just asking "did they win, yes or no?", their new measure traces which actions actually mattered — like checking how much of a group project's work really made it into the final report. This exposes wasted effort that a simple pass-or-fail score hides.

Why should you care? AI "agents" are being marketed right now as digital coworkers that can handle multi-step chores: booking travel, sorting email, managing orders. This test shows the teamwork is still shaky. The practical takeaway: keep a human in the loop on anything long or complicated, and don't expect a team of AI assistants to run a project unsupervised. The good news is the test is free and open, so progress can now be measured instead of guessed.