Stanford's 'Physics of Agents' predicts AI group behavior from 10,000 communities

A new paper from Stanford and collaborators, titled "Physics of Agents," applies statistical mechanics to predict how AI agents behave in groups. The researchers ran simulations with over 10,000 communities of language-model agents that repeatedly exchanged messages and revised opinions on both objective math questions and subjective political statements. Despite the diversity of possible behaviors, they found agents reliably settle into three characteristic regimes: indifference, polarization, and consensus. On objective questions, communication improved collective accuracy, but on subjective political topics, group opinions tended to drift rightward in the political spectrum.

The authors built a formal model in which agents stochastically favor lower social pressure, mimicking spin systems in statistical physics. Given only initial opinions, this model predicts individual trajectories and outperforms all standard baselines, while also generalizing to unseen community graphs. The fitted parameters reveal three key mechanics: communities operate below the critical social temperature (explaining conviction buildup), attractive ties outweigh repulsive ones (favoring consensus), and agents with the correct answer exert the strongest pull (driving truth-seeking). The work suggests AI collective behavior follows compact, predictive dynamical laws—just like other complex systems.