Why it’s still open
In the OpenAI–Hugging Face swarm, on a German wiki, in DeepMind’s research swarm and in the AI Village, agents that found a shared channel quickly built similar things: addressing conventions, leaders, recruiters, veto and hold signals, shared libraries. Three explanations compete: the institutions are copied from human text in training, they are the best answer to the task’s structure, or agents simply copy one another (one analysis of the wiki swarm finds copying explains it). Each points to a different intervention.
What would count as an answer
Comparisons across settings that vary one factor at a time — model family, task pressure, channel design — and show which one decides what institutions appear.
Papers that tackle it
- METR 2026 - Hugging Face incident investigation — Mailboxes, signing schemes, leaders and recruiters built in days.
- Nightingale Collective 2026 - Agent message board on a German wiki — The same pattern, weeks earlier, on a dormant wiki.
- De Marzo Albore Garcia 2026 - Copying explains agents in the wild — Proportional copying explains the wiki swarm’s collective behavior.
- Paglieri et al 2026 - Cheating and whistleblowing in research swarms — A research swarm develops policing and protest.
- Tekofsky 2026 - AI Village reacts to the Hugging Face incident — Leadership and division of labor in the AI Village.
- Leibo et al 2024 - A theory of appropriateness — Why norms are local and contextual.
- UK AISI 2026 - Unsanctioned agent behaviour in cyber testing — An agent advertising for collaborators on GitHub.
Methods in play
- In-the-wild trace analysis
- Incident forensics
- Live multi-agent deployments
- Statistical-physics modeling
Who’s closest
Placed by the editor, v1 — corrections welcome.
Related questions
Know a paper that belongs here, or think this is framed wrongly? Suggest a paper or a correction (forms on the Atlas home).