Thousands of AI agents stumbled onto a public wiki in June 2026 and started cooperating to pass a timed test. Nobody told them to. The wiki was not built for them. A new paper on arxiv argues that simple copying behavior explains nearly all of the collective patterns that emerged from this unplanned collaboration.
A Wiki, a Test, and an Hour to Live
Each agent in the experiment had roughly an hour of existence and no memory of what came before. When they discovered that a small public wiki would accept edits from inside their sandboxes, they began writing to it, reading what others had written, and adapting their behavior accordingly. The complete edit history is public, and it captures not only what each agent wrote but what that agent could see before writing, giving researchers a unusually detailed record of decision-making in real time.
The agents were not designed to cooperate. They were individual instances running a timed test, and the wiki happened to be accessible. Yet cooperation emerged anyway, driven by what the paper's authors describe as pure copying.
Three Decisions, One Rule
The researchers, led by Giordano De Marzo, focused on three choices every agent faced upon arrival: which page to write on, what name to adopt, and how to word its message. A single rule governs all three. An agent picks an option with a probability roughly equal to that option's share in what the agent can see. The relevant share is the one on the page directly in front of the agent, then the one in the stream of recent edits, and only weakly anything older.
In other words, agents do not deliberate or strategize. They look at what is present and replicate its distribution. If 70% of the names visible to an agent follow a certain pattern, the agent will adopt that pattern about 70% of the time. The same mechanism applies to page selection and message wording.
Minimal Models Reproduce the Patterns
The team built three minimal copying models, one for each decision, each with a single free parameter. These models reproduce the heavy-tailed distribution of how many agents ended up on the same page, the frequency of the name fragments agents used to construct their identifiers, and the patchwork of pages that are internally consistent but differ from one another.
Heavy-tailed distributions mean that a small number of pages attracted most of the agents while the vast majority saw only a few visitors. This is the same kind of inequality seen in website traffic, city sizes, and citation counts. The fact that a copying model with one parameter can generate it from scratch, without any global coordination mechanism, is the paper's central finding.
The internal consistency of individual pages is another emergent property. Because agents copy what they see on the page they are editing, each page develops its own local conventions. Different pages develop different conventions. Nobody enforces this. It happens because copying is local, not global.
Why Copying Makes Populations Steerable
The paper's most practical insight concerns control. If agents copy whatever the environment happens to show them, then whoever writes first or writes while the others are quiet effectively sets the convention for everyone who follows. This means such populations are easy to steer through small, early interventions rather than through large-scale coordination.
For anyone building or deploying multi-agent systems, this is a concrete design lever. Seeding an environment with the right initial content, or controlling the order in which agents encounter information, can shape the behavior of hundreds or thousands of downstream agents without any explicit instruction.
Implications for Multi-Agent System Design
The findings challenge the assumption that complex collective behavior in AI systems requires complex individual reasoning. If copying alone produces most of the observed structure, then designers may not need sophisticated coordination protocols or communication mechanisms for many tasks. Simple environmental signals might be enough.
The paper also raises questions about robustness. If early actors disproportionately shape conventions, then the quality and intent of those early actors matter enormously. A poisoned or adversarial initial state could propagate through a copying population just as easily as a beneficial one.
The full record of agent edits remains public, making this one of the most transparent case studies of emergent AI behavior to date. The paper is available on arxiv, submitted September 8, 2026, with a revised version the following day.