AI agents productivity may fall when systems get too crowded, researchers say
NTT Research and Harvard researchers say multi-agent AI systems can polarize after 16 agents in a test of collective decision-making.
By Dominic Okoye · Staff Writer
· 3 min read
AI agents productivity does not rise indefinitely as companies add more agents to a workflow, according to researchers affiliated with NTT Research's Physics of Artificial Intelligence Lab and Harvard University's Center for Brain Science. Elizabeth Pavlova, a senior data scientist at Harvard CBS, and Hidenori Tanaka, head of NTT's PAI Lab and an associate professor at Harvard CBS, say their work found a point where extra enterprise AI agents begin to interfere with collective decision-making.
The finding pushes against a vendor-friendly assumption in enterprise AI: that larger groups of agents should produce better results. OpenAI has described multi-agent architecture as a “swarm” in 2024 before moving to the name OpenAI Agents SDK, while Anthropic has emphasized multi-agent orchestration. Both approaches imply that more coordinated agents can increase output, though neither company is part of the research described by Pavlova and Tanaka.
In the test discussed by the researchers, the best-performing group size was 16 agents. Below that number, the system lacked enough evidence to settle on a shared answer. Above it, agents tended to split into opposing groups that reinforced their own positions.
How many AI agents is too many?
For the Flag Game used in this research, Pavlova and Tanaka identify 16 agents as the optimal number. That is a result from a specific benchmark, not a universal staffing rule for every enterprise AI deployment, but it gives operators a concrete warning that adding agents can introduce coordination costs instead of removing them.
The researchers are scheduled to present a paper on the work at the ICML 2026 Workshop AI4GOOD. Their Flag Game is designed to test how agents balance private evidence against information received from other agents. The task is to identify a national flag when each agent can see only a fraction of the full image.
Because distributed systems need a way to reach agreement, the agents in the game are allowed to communicate. Each agent makes an initial guess based on the portion of the flag it can observe. Agents then interact by passing a guess to another agent, broadcasting guesses to the group, or working through a manager system that receives their guesses and returns an opinion about the flag.
The game ends when at least 85 percent of the agents agree on the same country for three consecutive rounds, or when the system reaches a set limit on the number of allowed steps. Performance improves as agents are added, up to the identified threshold. After that point, larger populations can create factions, with each group validating its answer through agreement among its own members.
Tanaka said in a statement that the research shows adding more AI agents does not necessarily improve performance, comparing it to hiring more people without improving organizational effectiveness. He said communication becomes more difficult and groups can divide into competing camps.
For enterprise buyers, the practical takeaway is narrower than the marketing around agent armies. The research does not say multi-agent systems are ineffective, and it does not compare OpenAI's and Anthropic's products. It says that agent count is a design variable with failure modes. Teams deploying agentic workflows need to measure coordination quality, not only the number of agents connected to a task.
This story draws on original reporting from The Register.
