A September 30 arXiv preprint reports that teams in which each person used a separate AI agent produced worse group results than one agent handling everyone's requests in four simulated settings. The scenarios involved shared budgets, a software release queue, clinic appointments and group bookings; the study did not test agents in live workplaces.
The comparison changes who can see the whole task. In the coordinator setup, one agent received every user's request. In the team setups, each user had a separate agent, with either no peer messaging or a channel to contact other agents. The researchers tested up to five models across 77 scenarios, with model coverage and reasoning settings varying by environment.
Group bookings showed how communication can fail at the point of action. Across 25 personal-assistant scenarios and four models, the coordinator honored an explicit user's request about twice as often as peer-to-peer teams. Among team episodes with a committed order or booking, the request had been sent to the agent that made the commitment in 94–99% of cases, but it was available to that agent when it acted only 47–67% of the time. Teams kept the request less often when it was unavailable at the decision point.
A platform guard that blocked a commit until the agent read pending messages helped in reruns of previously failed team episodes: the user's request was honored after rerun in 73.1% of Claude Opus 5 episodes and 62.3% of Claude Sonnet 5 episodes. These are recovery rates for failed episodes, not overall success rates.
For products where separate agents change a shared resource, the study points to a concrete design question: can the agent committing an action see every pending request that should affect it? The results remain limited to constructed simulations, and the tested model mix differed across environments. The authors say they plan to release MAMUBench, a 74-scenario benchmark, but its linked repository was not public when checked.