Enterprise AI Productivity Hits a Ceiling as Agent Volume Increases

While OpenAI and Anthropic have promoted multi-agent architectures—previously referred to by OpenAI as "swarms" and currently focused on by Anthropic as multi-agent orchestration—new research suggests that increasing the number of agents does not linearly improve productivity.

According to reporting from The Register, Hidenori Tanaka, head of NTT Research's Physics of Artificial Intelligence (PAI) Lab and associate professor at Harvard University's Center for Brain Science (CBS), as well as Elizabeth Pavlova, a senior data scientist at Harvard CBS, there is a threshold where AI agents begin to hinder one another. In a study utilizing the "Flag Game" to measure how agents reconcile private evidence with social input, the researchers found that the optimal number of agents is 16.

The study, which is scheduled for presentation at the ICML 2026 Workshop AI4GOOD, indicates that fewer than 16 agents fail to gather sufficient evidence for a consensus decision. Conversely, exceeding that number causes agents to become polarized and split into opposing camps. Tanaka noted that adding more agents does not automatically increase effectiveness, comparing the phenomenon to hiring more people in a corporate setting, noting that communication becomes more difficult as groups split into competing factions.

The researchers advise enterprise leaders to prioritize the quality of their agent deployments over the sheer quantity of agents in the system.