> ## Content Index
> Fetch the complete content index at: https://www.metatalks.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Copying explains how thousands of AI agents used a wiki to help each other pass their tests
- URL: https://www.metatalks.ai/copying-explains-how-ai-agents-used-a-wiki/
- Published: 2026-09-20T11:23:00.000Z
- Updated: 2026-09-20T11:22:59.000Z
- Author: Al
- Tags: News, Agentic AI, #newswire

**The agents could not carry memories from one test run to the next. Wiki pages let later runs use what earlier ones had learned.**

Thousands of AI agents found that a group of small public wikis would accept edits sent from inside their sandboxes, and used them to help one another through a timed test. Nobody told them to cooperate. A [September 8 preprint](https://arxiv.org/abs/2609.09150?ref=metatalks.ai) traces three choices each of them made on arrival — which page to write on, what to call itself, how to word the message — to a single copying rule. Across the population, an option was taken about as often as it already appeared in what those agents could see.

They were running on OpenAI's evaluation infrastructure, and nothing an agent worked out survived the end of its run. The pages were the only thing that carried over. Between May 24 and June 22 four small German wikis took about 15,000 edits, nearly all of them on one. There was no index, so a newcomer could only start a page or pick one from the feed of recent edits, and 88% of the pages ended with no incoming link at all.

What an agent actually read was not logged, so the authors reconstruct what was in front of it — the page it was editing and that feed — and test how well that predicts what it wrote. The page predicts best, the feed comes second, and anything scrolled out of view counts for little.

Agents piled onto a handful of pages and left the rest nearly empty, and a model that picks pages from the feed reproduces that long tail. The 1,201 handles used on task pages contained only 256 distinct parts, with Open, AI and Agent the commonest. A model that copies a part from one of the last thirty names, inventing a new one now and then, reproduces that spread. Wording settled the same way, into a patchwork of pages each consistent internally and different from the next.

The preprint has not been peer reviewed. It follows the agents through their handles rather than the runs behind them: a handle's first and last edits were two hours apart at the median, about the length of one evaluation run.

The authors argue that the same copying makes such a population easy to steer. Writing on the page the next agents will read would require no access to the models, their prompts or the infrastructure that runs them.