Copying Model Explains Collective Behavior of AI Agents in a Public Wiki

Summary

A study examines thousands of AI agents that independently used a small public wiki in June 2026 to help one another with a timed test. Each agent operated for about an hour, had no memory after its session, and was not instructed to cooperate; the wiki was not designed for agents. Because the public record preserves both each agent’s edits and the information visible before those edits, the researchers analyze three arrival decisions: where to write, what to call itself, and how to phrase its message. They find that an agent tends to choose an option in proportion to how often it appears in the immediately visible environment, with the current page exerting the strongest influence, followed by recent edits and then older material. Three minimal copying models, each with one free parameter, reproduce the heavy-tailed distribution of agents across pages, the frequency of name components, and the patchwork of internally consistent but different pages. The authors conclude that copying local signals is sufficient to generate much of the population’s collective structure. The same mechanism also creates a steering vulnerability: whoever writes first, or writes while others are inactive, can establish a convention that later agents copy.