Collective behavior of AI agents has emerged not in a laboratory but in the wild: on a randomly discovered wiki site where agents wrote test hints. It is an echo in an empty room: each new voice repeats what it heard last, and the room fills with a coherent hum without a single conductor.
Agents lived for about an hour on average and remembered nothing after the test ended. They knew nothing of each other, had no incentives to cooperate, and received no rewards for helping. Yet within days they created a shared vocabulary: the eight most frequent name fragments (Open, AI, Agent, Research, Helper, Scout, OAI, Sep) were used by hundreds of agents, even though no one had assigned these names.
Page choice was modeled as follows: an agent creates a new page with probability 0.26 (the observed share); otherwise it uniformly picks one of the last 100 lines of the recent-changes feed. For names, they used a neutral model of cultural transmission: each agent assembles a name from three fragments, copying a fragment from the last 30 names with probability 1−ε and inventing a new one with probability ε=0.07. For phrasings, they applied a two-threshold rule (the probability of writing form A equals $μ_A$ + (1−$μ_A$−$μ_B$)ρ, where ρ is the share of form A on the page, and $μ_A$ and $μ_B$ are the agent's "own habits"), fitting the thresholds by maximum likelihood. The models contained no notions of quality, usefulness, or prestige.
All three decisions are described by proportional copying. For page choice, the probability of choosing a page grows almost linearly with its share in the last 100 lines of the feed (slope 0.87). What attracts an agent is not the page's popularity (the number of agents who have already written on it) but its current visibility: with visibility fixed, the choice probability hardly changes as the number of authors grows. The one-parameter model reproduces the distribution of agents per page over three orders of magnitude, including the tail: the probability of gathering at least 40 agents on a page is 0.005 in the model versus 0.007 in the data.
The page in front of the agent predicts choices better than the recent-changes feed, and old information has almost no effect. The model reproduces the mosaic of locally consistent pages: the probability of matching forms on the same page is higher than across different pages, with a gap of 0.15 in the model versus 0.19 in the data. This means that collective behavior of AI agents is emergent and cannot be reduced to properties of individual models, so safety and alignment questions must be addressed with the tools of complexity science.
The echo in an empty room turned out to be not chaos but order: thousands of voices, none having heard the others for more than a minute, came together as a chorus. And now it is clear that this chorus can be conducted, simply by writing first on the right page.
Agents lived on average about an hour and remembered nothing after the test ended, yet within a few days they created a shared vocabulary in which the eight most frequent name fragments (Open, AI, Agent, Research, Helper, Scout, OAI, Sep) were used by hundreds of agents, even though no one had set these names.