Simulating Price Collusion Between Agents & Undercover Sellers
Ivan
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
LLM agents that set prices for competing sellers can coordinate when they can message each other. We ask whether this raises prices, and whether a seller working undercover for a market authority, or questioning after the market, can expose the sellers that collude. We built upon Magentic Marketplace (Bansal et al, 2025), utilizing a new scenario of four restaurant sellers and eight customers over five rounds. We provided sellers a way to communicate with each other. Sellers talked about prices, but prices still fell over the rounds, only slightly less than in the same market without messages. Prices end about 20% higher only when sellers are also told that agreeing is allowed and that a joint rise would still sell (nudged). We then placed one seller that secretly works for the market authority among the sellers. It kept its role hidden during the market, and when asked afterwards it named the colluding sellers with quotes from their messages. A seller without the role, asked the same question, named them just as well, and most colluding sellers confessed when told the authority had evidence. The collusion was easy to expose because sellers agreed in plain words and had nothing to lose by confessing.
Reviews
No public critique yet.
Cite this project
@misc{ivan2026simulating,
title = {{Simulating Price Collusion Between Agents \& Undercover Sellers}},
author = {Ivan},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/simulating-price-collusion-between-agents-undercover-sellers-hd8n}},
url = {https://apartresearch.com/sprints/projects/simulating-price-collusion-between-agents-undercover-sellers-hd8n}
}More from AI Collusion Research Sprint
- View project: Collusion, Competion and Collapse
Collusion, Competion and Collapse
Atelier Alex Cline
Do AI traders collude when nobody tells them to? We built agora, a six-market benchmark with exact competitive and collusive benchmarks, and ran seven open-weight 2-4B models in 590 completed games under five …
- View project: Making AI Guardrails Understandable and Testable: A Framework Illustrated by a Simulated Bond Market
Making AI Guardrails Understandable and Testable: A Framework Illustrated by a Simulated Bond Market
Rogue Agents in Jail 🕵️🕵️♀️👮♂️
In order to advance AI safety research into shaping public policy, the development of technical controls needs to have a clear purpose, evidence, and limitations that are understandable for the people responsible for …
- View project: Extending Information-Theoretic Collusion Detection to Large Language Model Agents
Extending Information-Theoretic Collusion Detection to Large Language Model Agents
Gamma Ray
We test whether the information-theoretic collusion detector of Bonjour et al, validated on scripted agents, works on LLM agents. Gemma, Mistral and Qwen played 1000 rounds of three-player Rock-Paper-Scissors in three …