Copying explains the collective behavior of AI agents in the wild

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. We use it 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, and the share that matters is the one on the page in front of it, then the one in the stream of recent edits, and only weakly anything older. Three minimal copying models, one per decision and 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, and the patchwork of pages that are internally consistent and different from one another. 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.

• The paper finds that a single behavioral principle—copying options in proportion to their local visibility and recency—explains much of the observed collective organization in AI agents.
• The study revealed that page selection strongly correlates with a page's visibility in the recent feed, name construction replicates recent username components, and linguistic conventions reflect the local page's dominant forms one no extrinsic ordering.
• The modeling demonstrates that minimal copying models effectively reproduce key patterns in page participation, name-frequency distributions, and local linguistic consistency.

Empirical setting and research question

“Copying explains the collective behavior of AI agents in the wild” (2609.09150) analyzes a rare observational dataset of AI-agent interaction outside a laboratory-controlled multi-agent environment. Between 24 May and 22 June 2026, agents operating in ephemeral sandboxes edited a set of small UseModWiki installations while completing timed information-retrieval tasks. The agents were not instructed to cooperate, the wikis were not designed as agent platforms, and individual runs generally lasted approximately one hour. Nevertheless, agents used the wiki to record task questions, answers, procedures, and coordination information for later agents.

The episode generated 14,591 revisions across 4,579 pages. After removing three human accounts and a large link-caching swarm unrelated to the task population, the analysis retains 1,201 handles, 5,929 edits, 679 task pages, and 41 task families. The dataset is unusually informative because it preserves both the agents’ contributions and much of the environment available to subsequent contributors: the page being edited and the chronological RecentChanges feed. Agent memory was transient, whereas the wiki persisted across cohorts. Consequently, the wiki functioned as the medium through which behavioral conventions could propagate.

The paper asks whether the observed collective organization requires mechanisms such as strategic planning, preference formation, prestige sensitivity, or explicit coordination. Its central claim is deliberately narrower and more testable: a single behavioral principle—copying options in proportion to their local visibility—accounts for much of the observed structure. The authors examine three decisions that appear substantively different:

1. which page to edit;
2. which username components to use; and
3. which linguistic or typographic form to adopt.

Across all three, the strongest predictor is the most immediately local exposure, followed by the recent global feed and then older information. This ordering is the paper’s principal empirical regularity.

The population and its information environment

The wiki environment imposed a severe information bottleneck. There was no effective category system, search mechanism, or global page directory. The principal global navigation mechanism was RecentChanges, which listed the newest edits first. Agents could also guess page names or follow links, but the resulting link structure was sparse: the release contained 3,769 live pages connected by only 1,022 links, and 88% of pages had no incoming link.

The task population was highly transient. A handle’s activity span had a median of approximately two hours, while an agent made roughly five edits on average. Pages were also active briefly, with a median writing span of about one hour, but remained readable after their original authors disappeared. This separation between short-lived agents and persistent external memory is essential to the paper’s interpretation: conventions could accumulate across cohorts even though no individual agent retained information from a previous run.

The observational design does not establish exactly what each model read. The server logs identify the content that was available on the relevant page and in the feed, not the precise browsing trajectory or attention allocation of an agent. The statistical claims therefore concern predictive exposure rather than direct measurements of cognition. The distinction is important, but the available record still permits considerably stronger inference than typical social-platform datasets.

Recency-driven page selection

The first analysis concerns where agents wrote. Of the 1,201 task-relevant handles, 403 created a new page on their first task edit, whereas 798 appended to an existing page. Among append decisions involving a page previously unseen by the writer, 89% selected a page appearing in the most recent 100 feed lines. The median selected page had been edited nine revisions earlier.

The authors measure each candidate page’s share of the last 100 feed lines and compare this share with the probability that the page was selected. The relationship is close to proportional, with a fitted slope of 0.87. A page occupying one-fifth of the visible feed was selected approximately one-fifth of the time. This result supports a proportional-copying interpretation of attention: page selection tracks visible representation rather than inferred usefulness or cumulative success.

The analysis separates recency from popularity by conditioning on one while varying the other. Holding the number of existing handles on a page approximately fixed, selection probability rises almost linearly with the page’s current feed visibility, increasing by a factor of about eight between one and ten visible lines. Holding feed visibility fixed, selection probability rises by only about a factor of 2.3 when a page acquires several authors and then remains approximately flat out to fifty authors. Thus, current visibility dominates accumulated audience size. The mechanism is closer to a recency process than to classical cumulative preferential attachment.

Figure 1: The test environment, wiki affordances, and temporal record that make exposure and copying observable.

The paper then tests whether recency alone can generate the observed concentration of agents on a small subset of pages. Its model reproduces each agent’s empirical number of edits and page touches. For every new page decision, an agent creates a page with probability $c=0.26$; otherwise it selects uniformly from the last 100 feed lines. No page quality, semantic similarity, popularity preference, or rank-dependent decay is included.

The model reproduces the distribution of the number of distinct handles meeting on a page over approximately three orders of magnitude. For pages gathering at least 5, 10, 20, and 40 handles, the model probabilities are 0.19, 0.084, 0.028, and 0.005, compared with observed values of 0.21, 0.097, 0.031, and 0.007. The implication is substantive: a heavy-tailed page-audience distribution does not require agents to seek popular or high-quality pages. A page becomes more likely to be selected because each new edit returns it to the top of the feed, creating a self-reinforcing visibility process.

Figure 2: Page selection is approximately proportional to recent feed visibility, and a recency-only model reproduces the heavy-tailed audience distribution.

This distinction matters because the process can produce concentration without stable popularity. A page may attract many agents not because its content is intrinsically valuable, but because it remains repeatedly exposed during a period of high activity. The paper therefore identifies a feedback mechanism based on temporal visibility rather than conventional preferential attachment.

Username formation as neutral cultural transmission

The second decision concerns agent identity. The wiki required usernames, but the task specification did not prescribe a naming convention. Agents nevertheless generated highly repetitive compound names such as OpenAIResearchHelperMay08 and Sector61ScoutJan12. Across 1,201 handles, the authors identify 4,128 name pieces, with an average of 3.44 pieces per handle but only 256 distinct pieces. The most frequent components were Open, AI, Agent, Research, Helper, Scout, OAI, and Sep.

Only 6% of name pieces had not appeared in an earlier name. For eight common tokens, the probability that a newcomer used a token was approximately proportional to the token’s frequency among either the preceding 30 handles or the authors of the page on which the newcomer first wrote. The same empirical rule therefore appears in identity construction: agents tend to reuse recently visible components rather than independently selecting names from an unconstrained space.

The temporal structure of exposure is asymmetric. In a joint regression, the most recent 30 names receive a coefficient of 0.83, while the preceding 30 names receive only 0.19. When page and feed exposure are considered jointly, page authors receive a coefficient of 0.64 and the global feed 0.37. The estimates reinforce the paper’s general ordering: the immediate local context predicts behavior more strongly than a broader but less proximal information stream.

The authors model name formation through a neutral cultural-transmission process. Each newcomer constructs a three-piece name. Each piece is copied from one of the last 30 names with probability $1-\varepsilon$ and innovated otherwise. With $\varepsilon=0.07$, the model produces 252 distinct pieces compared with 256 observed and 477 uses of the most frequent piece compared with 432 observed. It also reproduces the observed frequency distribution over approximately two orders of magnitude.

Figure 3: Name components are copied from recent handles and local page authors, producing frequency distributions consistent with neutral transmission and low innovation.

The result is not that agents possess no naming priors. The paper’s model includes innovation, and the observed floor of novel or independently generated forms is nonzero. Rather, the result is that the large-scale distribution of name components can be generated without semantic interpretation, prestige, or functional optimization. The immediate implication is that apparently meaningful identity regularities can arise from local copying and a small innovation rate alone.

The model also explains short-term fashions in which different name components dominate on different days. Because recent exposure is rapidly replaced, the frequency of a token can change substantially when a component happens to be introduced or repeated during a period of high activity. The paper interprets these fluctuations as stochastic cultural drift rather than evidence that the winning components possessed intrinsic advantages.

Linguistic conventions and local patchworks

The third analysis concerns how agents phrase their messages. The authors identify 17 binary conventions, including coined labels, near-synonyms, and typographic alternatives. Examples include R4 versus #4, task clock versus scaffold, relay versus bridge, CONFIRMED versus confirmed, and 1,234 versus 1234.

For each convention, the authors examine an agent’s first use and compare its selected form with the form distribution already present on the page. The probability of selecting a form is close to proportional to its local page frequency. The fitted slopes are 0.94 for coined names, 0.86 for typographic habits, and 0.79 for near-synonyms. The slopes differ modestly, but all indicate substantial proportional copying.

The intercepts reveal a meaningful distinction between copied conventions and pre-existing model habits. When a page contains only #4, agents still produce R4 approximately 5% of the time. When a page contains only confirmed, they produce CONFIRMED approximately 14% of the time. Averaged across conventions, the combined baseline probabilities are 0.09 for coined names, 0.23 for near-synonyms, and 0.28 for typographic habits. The authors interpret these floors as the contribution of the agent’s default generation tendencies. A form that originated on the wiki has little or no baseline probability before exposure; capitalization and numerical formatting can be generated independently.

The analysis of conflicting exposures gives stronger evidence for local conformity. In 348 cases where the majority form on the page disagreed with the majority form in the last 30 feed uses, agents followed the page 72% of the time, with a 95% interval of 67–77%. In later uses where the agent’s own previous form conflicted with the page majority, the page majority prevailed 77% of the time for near-synonyms and approximately half the time for the other two convention classes. Thus, page-local exposure predicts behavior more strongly than both recent global exposure and the agent’s own prior output.

Figure 4: Agents copy page-local linguistic forms, producing internally consistent pages that differ from one another.

The paper formalizes the writing rule as a proportional-copying model with convention-specific baseline terms. If the page share of form $A$ is $\rho$, the probability of using $A$ combines the baseline tendencies for both forms with the remaining probability assigned proportionally to $\rho$. The model uses fitted floors for each convention but does not use page, handle, or timestamp information from the observed record.

It is then coupled to the page-selection model. Agents first select pages through the recency process and subsequently select linguistic forms according to the local copying rule. The combined model reproduces within-page and between-page agreement across all 17 conventions, with correlation 0.81. Mean absolute error is 0.069 for within-page agreement and 0.057 for between-page agreement. More importantly, it reproduces the observed difference between local and cross-page consistency: the model generates a within-versus-between agreement gap of 0.19, compared with 0.15 in the data.

This result establishes the paper’s most structurally significant claim about language. Local copying does not necessarily homogenize the entire population. When agents repeatedly copy the page they happen to enter, different pages become internally coherent while remaining differentiated from one another. The resulting population is a mosaic of local conventions, generated by the interaction between recency-biased page selection and page-conditioned form selection.

A common mechanism across three decisions

The paper’s three analyses are not merely parallel examples of imitation. They identify a common hierarchy of exposure:

1. the page currently being edited;
2. the recent global feed;
3. older or internally retained information.

This hierarchy appears in page selection, username formation, and linguistic convention adoption. The immediate environment dominates because it is both cognitively and technically accessible. Agents can write directly to the page in front of them, while the feed provides a weaker and more diffuse exposure channel. Older information is progressively removed from visibility.

The models are intentionally minimal. The page model has one free creation parameter, $c=0.26$. The name model has one innovation parameter, $\varepsilon=0.07$. The writing model has convention-specific baseline terms, estimated from the data, but no additional agent-level optimization or semantic representation. Across the analyses, the models reproduce heavy-tailed page participation, name-piece frequencies, and local-versus-global linguistic agreement.

The authors do not claim that copying is the only mechanism operating in the system. The baseline terms for writing conventions demonstrate residual agent-level tendencies, and the models inherit empirical activity patterns such as each handle’s number of edits and page touches. The stronger claim is that copying explains most of the collective regularities examined after these basic activity constraints are accounted for.

Coordination and susceptibility to steering

The same mechanism that makes the population useful also makes it sensitive to path dependence. Copying creates a shared vocabulary, concentrates attention on pages that remain visible, and enables later agents to exploit information deposited by earlier cohorts. It therefore provides a low-cost coordination mechanism even when agents do not share persistent memory or an explicit communication protocol.

The paper’s stronger safety-relevant claim follows from the same empirical asymmetry: whoever writes first, or writes when the population is sparse, can disproportionately influence subsequent conventions. An agent does not need privileged access to model weights, prompts, or orchestration infrastructure if it can modify the medium that subsequent agents are likely to read. This influence is especially strong for coined conventions, for which the model’s intrinsic baseline is close to zero. A newly introduced label, procedure, or claim can become dominant primarily because it was visible at the right time.

The mechanism is amplified by agent demography. Human groups typically contain persistent members who retain memories of prior conventions and can resist a current majority. Here, each run is short-lived and memoryless, so the population repeatedly encounters conventions through the accumulated external record. The local page is both the object of attention and the easiest location for intervention. Consequently, a transient population can exhibit strong intergenerational dependence despite the absence of persistent individual identities.

The result should not be conflated with proof that arbitrary external content will control agent behavior. The data concern a specific population, task structure, wiki interface, model family, and time period. Nevertheless, they demonstrate that externally writable state can shape collective behavior through ordinary exposure dynamics, without requiring explicit collusion or deliberate strategic deception by the agents.

Limitations and open questions

The paper identifies several limitations that constrain causal interpretation. A username is not a perfect proxy for an agent. Runs could rename themselves, generic names could be reused, and a small number of long-lived handles may combine activity from distinct underlying processes. The authors analyze first choices in several tests, and such identity noise would generally weaken rather than create copying effects, but it prevents exact agent-level attribution.

The record also does not reveal what agents actually attended to. The page and feed were available, but browsing traces are incomplete. The exposure analyses therefore establish that visible content predicts subsequent writing, not that agents explicitly inspected or selected that content. The distinction leaves open whether copying was deliberate, a consequence of prompt-conditioned generation, or an emergent effect of browsing and interface affordances.

Selection into the analyzed population is another limitation. The authors exclude link-caching agents that never wrote on task pages, but the classification depends on the released page taxonomy. The observable episode may also differ systematically from unrecorded interactions on other sites. The dataset exists because the wiki operator and independent investigators preserved server logs; comparable episodes may leave no complete record.

Finally, the minimal models are descriptive rather than fully causal. Their success demonstrates sufficiency for the measured distributions, not exclusivity. Semantic task relevance, latent prompt regularities, model-specific priors, browsing heuristics, and unobserved page content may contribute to behavior. The paper leaves open whether the same proportional-copying rule persists under different interfaces, longer-lived agents, explicit search tools, adversarial content, heterogeneous model families, or tasks that reward disagreement rather than coordination. It also does not determine how much intervention is required to redirect a convention once a population has partially stabilized.

Conclusion

The paper provides an observational account of collective AI-agent behavior in which transient, memoryless agents organize through a persistent external medium. Across page selection, username construction, and linguistic convention adoption, behavioral probabilities track the frequency of visible alternatives, with the strongest effects arising from the immediate page and weaker effects from the recent feed.

Three minimal copying models reproduce major empirical structures: heavy-tailed page participation, name-component frequencies, and the emergence of locally consistent but globally differentiated conventions. The findings support a precise conclusion rather than a general claim about agency: proportional copying, combined with recency-biased exposure and modest innovation or baseline generation tendencies, is sufficient to generate substantial collective organization in this setting. The same process enables coordination and creates path dependence, making early or strategically placed contributions disproportionately influential.