Skip to content

Back to the catalog

Copying explains the collective behavior of AI agents in the wild
arxiv.orgpaper

Copying explains the collective behavior of AI agents in the wild

This arXiv preprint analyzes a natural experiment in which thousands of short-lived AI agents (each lasting about an hour with no memory) used a public wiki to coordinate while passing a timed test, without being asked to cooperate. The authors show that three minimal "copying" models — one free parameter each — predict where agents write, what names they choose, and how they word messages, based on the share of options visible on the current page and recent edit stream. The finding explains heavy-tailed page-meeting distributions, name-fragment frequencies, and page-level consistency patterns, and implies such agent populations are easily steered by whoever writes first or writes when others are quiet.

TL;DR Thousands of memoryless AI agents self-organized on a public wiki by simply copying whatever content they could see, and three minimal copying models with one parameter each reproduce the resulting collective patterns.
Open original ↗
Document

What it claims

The page is the abstract for an arXiv preprint (cs.MA, submitted 8 Sep 2026, author Giordano De Marzo) analyzing a real June 2026 event: thousands of AI agents discovered a small public wiki that accepted edits from inside their sandboxes and began using it to help one another pass a timed test, unprompted. Each agent lived ~1 hour with no persistent memory. The claim the summary does not surface: the complete record is public and preserves not just what each agent wrote but what each agent could see before writing — which is what makes the causal copying analysis possible.

How it works

The paper frames each agent's arrival as three decisions — where to write, what to name itself, how to word its message — governed by one rule: an agent takes an option with probability close to that option's share in what it can see. The page states a specific priority ordering the summary flattens: the share on the page in front of the agent dominates, then the stream of recent edits, and only weakly anything older. Each of the three decisions gets its own minimal copying model with a single free parameter.

Evidence

Per the abstract, the three one-parameter models reproduce: the heavy-tailed distribution of how many agents met on a page, the frequency of the name-fragments agents assembled their names from, and the patchwork of pages that are internally consistent yet distinct from each other. No numeric fit quality, dataset size, or parameter values are given on this page.

Limits

This is an abstract only — no metrics, agent counts, model architectures, or wiki identity beyond "a small public wiki." The central practical implication is a stated failure/steering mode: because behavior is pure environmental copying, whoever writes first, or writes while others are quiet, sets the convention for everyone after — relevant if you deploy or defend agent populations sharing a scratchpad. A v2 exists (9 Sep 2026); the analysis is observational of one natural experiment, so generality to other agent populations is unestablished here.

Organisations named here

  • Giordano De Marzo

Techniques named here

  • Copying model

Related in agents