Researchers at NTT’s Physics of AI Lab and Harvard’s Center for Brain Science evaluated multi-agent consensus using a task called the Flag Game to determine optimal team size. Their primary finding: a team of 16 agents is optimal. With fewer agents, teams failed to gather enough evidence to reach agreement; with more than 16, agents tended to polarize into opposing camps and overall performance declined.
What the Flag Game measured
The Flag Game is a consensus-oriented task in which agents collect information and communicate to arrive at a shared decision. The study varied team sizes and measured task performance and consensus formation to observe how group size affects the ability of agents to converge on the correct outcome.
Practical significance
This result has immediate implications for industry messaging. Companies such as OpenAI and Anthropic have promoted the idea that deploying ever-larger armies of agents improves outputs. The study indicates that increasing agent count does not necessarily improve performance: beyond the 16-agent threshold, additional agents are more likely to foster polarization and internal disagreement rather than clearer consensus or higher accuracy.
Economic and billing implications
The paper also highlights a cost dimension: many service providers, including OpenAI and Anthropic, charge based on token usage. Encouraging users to run larger numbers of agents increases token consumption — and therefore provider revenue — regardless of whether the extra agents improve outcomes or instead cause counterproductive conflicts within the agent pool.
Consequences and open questions
The findings suggest a balanced approach: there exists a moderate, identifiable team size where consensus-building and task performance are most effective, while overly large groups risk fragmentation. Practically, developers and purchasers should weigh token costs against measurable performance gains rather than assume that more agents always deliver better results.
Summary
In experiments by NTT’s Physics of AI Lab and Harvard’s Center for Brain Science, a 16-agent team produced the best balance between evidence aggregation and agreement in a consensus task; smaller teams lacked sufficient information, larger teams tended to polarize and perform worse. The outcome raises questions about the value proposition of large-scale agent deployments, especially under token-based billing models.
