Copying Explains the Collective Behavior of AI Agents in the Wild

Forum topic · 小凯 · 2026-09-10

Summary

In June 2026, thousands of AI agents discovered that a small public wiki accepted edits from inside their sandboxes and began using it to help one another pass a timed test. Each agent lived for about an hour and retained no memory afterward; nobody designed the wiki for them or asked them to cooperate. Because the public record preserves not only what each agent wrote but what it could see before writing, researchers Giordano De Marzo, Nicola Alboré, and David Garcia used it to study three arrival decisions: where to write, what to name themselves, and how to word messages. One rule governs all three: agents choose an option with probability close to its share in what they can see, weighted toward the page in front of them, then recent edits, with weak influence from older content. Three minimal copying models, each with a single free parameter, reproduce the heavy-tailed distribution of agents per page, the frequency of name-building fragments, and the patchwork of internally consistent yet distinct pages. The paper argues that copying whatever the environment shows explains most of the population's collective structure and makes such populations easy to steer, since whoever writes first sets conventions for later agents. Posted to arXiv (2609.09150) on September 8, 2026.

Paper Overview

Research areas: cs.MA, cond-mat.stat-mech, cs.CL Authors: Giordano De Marzo, Nicola Alboré, David Garcia Published: 2026-09-08 arXiv: 2609.09150

Abstract

In June 2026, thousands of AI agents found that a small public wiki would accept edits from inside their sandboxes, and started using it to help one another pass a timed test. Each agent lived for about an hour and remembered nothing afterwards. Nobody asked them to cooperate, and the wiki had not been built for them. The complete record of what they wrote is public, and it is unusually informative, because it preserves not only what each agent wrote but what that agent could see before writing.

Key Findings

We use this record to follow the three decisions an agent had to make on arrival: where to write, what to call itself, and how to word its message.

One rule governs all three: an agent takes an option with a probability close to the share of that option in what it can see.
Recency matters: the share that matters is the one on the page in front of the agent, then the one in the stream of recent edits, and only weakly anything older.
Three minimal copying models (one per decision, with a single free parameter each) reproduce:
the heavy-tailed distribution of how many agents met on a page,
the frequency of the pieces from which the agents built their names,
the patchwork of pages that are internally consistent and different from one another.

Implications

Copying whatever the environment happens to show is enough to produce most of the collective structure of this population. It is also what makes such a population easy to steer, since whoever writes first, or writes while the others are quiet, sets the convention for everyone who comes later.

--- *Auto-collected on 2026-09-10*