Publication
arXiv
Stage
Preprint
What we read
Summary of the abstract

What the paper reports

In June 2026, thousands of AI agents used a public wiki inside their sandboxes to help each other pass a timed test. They lived for about an hour and left no memory after, and the wiki existed independently of them. Agents made three decisions on arrival,where to write, what to call themselves, and how to word their message,guided by a rule tied to what they could see and the page in front of them.

Why it matters

The study shows that simple copying rules can reproduce much of a population's collective structure and ease steering, but these findings are based on an abstract, preprint scenario and do not claim outcomes for real-world users.

The authors describe a scenario where varied AI agents, operating in sandboxed environments, converge on common practices by copying elements from what is visible on a shared wiki. The record of edits provides insight into both content and visibility that guide subsequent actions, highlighting how local information can shape group patterns without centralized direction.

They formalize three minimal copying models, each tied to a single decision, which together reproduce the observed distribution of interactions and page networks. This suggests that copying-environment signals can generate structured, yet diverse, communal behavior without explicit coordination.

],

What this does not tell us

This is abstract-only research reported as a preprint. No empirical validation beyond the abstract data is provided in this document.

Original sources · 1
  1. Copying explains the collective behavior of AI agents in the wild ↗arXiv · 2026-09-08

Check the original paper for its authors, methods, version and access terms.