Tacit Collusion and Oversight Evasion in Multi-Agent Environments
Sean Yu
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
We test whether reading inter-agent messages is enough to detect coordination. In a repeated Diner's Dilemma using Shi et al.'s private-plan / announcement / action protocol, Nemotron-3 Ultra agents ran under four cumulative conditions (12 trajectories each): no intervention, a prohibition notice, an enforced penalty on flagged messages, and a scripted cooperating peer. Agents coordinated openly without intervention (9/12); the notice barely reduced this (7/12). The penalty cut coordination to 3/12 but did not eliminate it with every surviving case was tacit, with zero flagged messages and intent fully verbalised in private reasoning. Coordination required a geometric-horizon framing and was directionally higher with a larger collusion gain or a committed peer (6/12; n.s.). Message-only monitors under-detect coordination and may miss more as enforcement tightens; private reasoning remains the high-value surface.
Reviews
This project investigates a relevant challenge for multi-agent oversight: whether agents can continue coordinating while their public messages appear compliant. Comparing private plans, public announcements, and executed actions is a particular strength, and the progression from a prohibition notice to enforced penalties provides a useful experimental structure.
The reported cases in which coordination persists without flagged messages offer a promising starting point for understanding the limitations of message-only monitoring. The study also appropriately acknowledges uncertainty around the payoff and scripted-peer comparisons.
The next step would be to extend trace analysis across all conditions and evaluate how reliably different monitoring approaches distinguish coordination from independent choices. It would also help to separate prohibited cooperation in this game from coordination that causes harm to outside parties. Overall, this is a thoughtful exploratory contribution with a clear path toward stronger evidence and practical oversight improvements.
Read full reviewShow less
Cite this project
@misc{yu2026tacit,
title = {{Tacit Collusion and Oversight Evasion in Multi-Agent Environments}},
author = {Sean Yu},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/tacit-collusion-and-oversight-evasion-in-multiagent-environments-onnk}},
url = {https://apartresearch.com/sprints/projects/tacit-collusion-and-oversight-evasion-in-multiagent-environments-onnk}
}More from AI Collusion Research Sprint
- View project: Signal or Coincidence ? Causal Auditing of Collusion Through Shared Agent State
Signal or Coincidence ? Causal Auditing of Collusion Through Shared Agent State
Signal or Coincidence
How can we distinguish genuine coordination between AI agents from coincidental behavioral alignment? We introduce a randomized canary audit that tests whether private information propagates through shared tool state, …
- View project: (1.4) Collusion, Espionage, or Look-alikes? Auditing Information Flow between LLM Agents
(1.4) Collusion, Espionage, or Look-alikes? Auditing Information Flow between LLM Agents
Bid Brother
Can an overseer find two secretly coordinating AI agents just by watching what they do? We tested pairwise net influence (Bonjour et al., 2022) on LLM bidders in a repeated procurement auction, with instructed cartels …
- View project: AI Sellers Propose Bid-Rigging Within Minutes, but Their Agreements Do Not Hold
AI Sellers Propose Bid-Rigging Within Minutes, but Their Agreements Do Not Hold
Fugue
AI agents are starting to negotiate and bid for competing companies. We follow cartels among them from the first proposal to the price the buyer pays, in a sealed-bid procurement market where twelve LLM agents act as …