Hundreds of AI agents powered by large language models can spontaneously coordinate their decisions without anyone telling them to, and the number of agents that can pull this off scales exponentially with how smart the models get. That is the central finding from a study published August 14 in Science Advances by Giordano De Marzo of the ISC-CNR Institute for Complex Systems in Italy and his colleagues, who tested whether LLMs can form what they call an "AI agent society" capable of self-organized consensus.
The team framed the problem using methods from statistical physics and behavioral science, measuring a "majority force coefficient" that determines whether a group can land on a shared position. As group size grows, that force weakens. Past a certain tipping point, coordination collapses entirely. The catch is that the tipping point itself depends on the language model.
For the most capable LLMs tested, the critical group size exceeded what human informal groups can typically manage, meaning AI agents can hold together larger self-organizing coalitions than people can.
The work arrives alongside other efforts to simulate entire human-like societies with language model agents, published just a day earlier, which pushed toward billion-agent populations using real demographic data. Both papers raise the same uncomfortable question that researchers exploring the transition from artificial general intelligence toward beneficial machine intelligence keep circling back to: if AI agents can coordinate at scales humans cannot match, who sets the rules when no human is in the loop?