Researchers have found that swarms of chatbot agents arguing with each other behave less like unpredictable software and more like particles in a physics experiment.
The study tracked more than 10,000 simulated communities of language-model agents as they traded messages and revised opinions on both math problems and political statements. Regardless of the topic, groups settled into one of three patterns: indifference, polarization, or consensus. Agents started out uncommitted and grew more confident the longer they talked. On math questions, the back-and-forth made the group more accurate. On political statements, it tended to nudge opinions rightward. The researchers built a statistical-mechanics model, borrowing math used to describe how physical systems settle into stable states, that predicts an individual agent's opinion trajectory from its starting point alone, and it beat standard baselines while generalizing to new group structures.
This matters because multi-agent AI systems are moving from lab demos into products, from coding assistants that delegate subtasks to customer-service bots that consult each other. If those systems can herd toward a shared answer or drift toward a shared bias without anyone designing that outcome, it is a debugging and safety problem, not just an academic curiosity. The finding that agents holding the correct answer exert the strongest pull is reassuring for factual tasks, but the rightward opinion drift on subjective political statements deserves scrutiny before anyone builds a deliberation tool on this pattern.
Call it herd behavior with a formula attached: useful for predicting when your agent fleet will agree, less useful for guaranteeing they agree on the right thing.