Researchers at Stanford University have released a study titled Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams, submitted to arXiv on September 30, 2026. The paper examines how teams of AI agents perform when sharing limited resources, such as API-token budgets, clinic scheduling, and code-merge queues.
The team evaluated five advanced models across 77 scenarios using a new benchmarking framework called MAMUBench. The results indicate that single-agent coordinators consistently outperform multi-agent teams. While single-agent setups achieved 32–64% of optimal outcomes, peer-to-peer teams reached only 12–30%.
Silent teams performed the worst, with success rates dropping as low as 2–7% in some cases. Even when communication channels were provided, agents frequently stalled or overrode the actions of their peers. The study notes that performance participation declined from 66–82% with four agents to 10–25% with 16 agents.
The researchers conclude that the perceived benefits of multi-agent strategies often stem from increased computational resources rather than effective collaboration. The paper suggests that decentralization can exacerbate conflicts between agents, and provides diagnostic details on failure modes along with proposed fixes for specific environments.
