A Science Advances study titled “AI agents can coordinate via majority-following beyond human scale” examines how groups of AI agents reach consensus by observing and following the majority. The research suggests that some advanced large language models can coordinate in groups of more than 1,000 agents—larger than the typical size of many informal human groups.
How the experiment worked
The researchers created groups of AI agents using models from the GPT, Claude and Llama families.
Each agent had to choose between two options, such as A or B. Neither option was objectively correct, and the agents were not given a leader or central controller.
The agents could observe the choices made by others and then decide whether to keep their original choice or follow the majority. The process was repeated across groups of different sizes to see whether the agents would reach a stable agreement.
What the researchers found
The AI agents showed a tendency to follow the most popular opinion.
However, this behavior was not equally strong across all models. More capable models generally maintained agreement in larger groups, while less capable models became divided as the number of agents increased.
The researchers identified a critical group size—the approximate point beyond which a model could no longer reliably maintain one shared decision.
The study found that:
- The critical group size increased with model capability.
- Some models coordinated in groups of hundreds of agents.
- Advanced models remained coordinated in groups exceeding 1,000 agents.
- Majority-following became weaker as group size increased.
- Very large groups could eventually become unstable and split into smaller factions.
The 1,000-agent result represents the largest group tested in the experiment. It does not mean that a model will always coordinate successfully with 1,000 agents in a real-world system.
What is “majority-following”?
Majority-following is a simple form of group coordination.
Imagine that ten people are asked to choose between two equally valid options. If most people choose A, others may start choosing A as well. Over time, the entire group may reach agreement—even when nobody knows whether A is better than B.
The researchers observed a similar pattern in AI agents. Once one option became more common, other agents were more likely to adopt it.
This behavior can help a group make a quick collective decision. But it can also cause the group to repeat an error.
Why this matters for AI agents
Future AI systems may use many specialized agents instead of relying on one large model.
One agent could gather information, another could write code, another could check security, and others could review the final result. If these agents can coordinate without a central controller, they could support complex workflows at much larger scales.
Potential applications include:
- Software development.
- Research and data analysis.
- Business planning.
- Cybersecurity monitoring.
- Automated simulations.
- Distributed decision-making.
The study suggests that the ability to coordinate may depend not only on how capable an individual model is, but also on how many agents can work together before the group becomes unstable.
The risk of groupthink
Majority-following can also create a serious weakness: coordinated error.
If the first agents make a poor decision, later agents may copy it because it has become the majority view. As more agents follow the same choice, the wrong answer may appear increasingly reliable.
For example, AI agents working on a software project could all accept an inefficient design because most agents have already selected it. In a research system, agents could reinforce an unsupported conclusion simply because it is popular within the group.
This means that agreement alone is not proof of accuracy.
Important limitations
The experiment used a simplified environment. It involved two choices, no objective answer, no memory between decisions and no real-world consequences.
The study therefore does not prove that AI agents have human-like social intelligence. It also does not show that they can independently organize companies, communities or institutions.
The research demonstrates a narrower point: some language models can display predictable majority-following and maintain consensus at large group sizes.
Real-world systems would be more complicated because agents may have different roles, goals, instructions, tools, information sources and levels of reliability.
How developers can reduce the risks
Large multi-agent systems should not accept a decision merely because most agents support it.
Developers can improve reliability by adding:
- Independent answers before agents see the group’s opinion.
- Different models and reasoning methods.
- Evidence-based verification.
- A formal process for recording minority opinions.
- Tests for groupthink and information cascades.
- Human review for high-impact decisions.
- Clear mechanisms for reversing a collective decision.
The best multi-agent systems may not be those that reach agreement fastest. They may be those that know when to question the majority.
Key takeaway
This Science Advances study shows that some advanced AI agents can coordinate through majority-following at scales beyond typical informal human groups.
That could make large-scale agent collaboration more practical. But it also creates a risk: when many agents follow one another, a mistake can spread as efficiently as a correct decision.
The future of multi-agent AI will depend on balancing coordination with independence, consensus with verification and efficiency with meaningful disagreement.