Decision Theory and AI Collusion: How Decision Rules Shape Agentic Collusion
Briant Moreno Abad, Juan Enrique Stücker Martinez · Team moo-nopoly
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
LLM pricing agents can reach above-competitive prices without communicating, but the role of decision-theoretic reasoning remains unclear. We test whether prompting evidential (EDT) or causal (CDT) decision theory, both paired with the premise that rivals are copies of the same model, changes coordination relative to the LLMCartel baseline, with two or three sellers over 10 rounds and four seeds. With three sellers, EDT yields the highest mean profit gain and the widest variation across seeds, and is the only arm with a run near the joint optimum. With two sellers, the control drifts above Nash on its own, and EDT either coordinates from the first rounds or ends at or below Nash. Once collusion forms, removing one agent's EDT paragraph rarely breaks it, while a best-response bot breaks it only after two to four rounds. Traces show CDT agents falling back on twin reasoning, and EDT coordination persisting only when the symmetric argument survives the first low price. Because EDT and CDT share the same-model premise, the effects cannot be attributed to the evidential rule alone.
Reviews
I liked that this paper asks a focused question and tests what happens after coordination forms, rather than only reporting high prices. Removing the EDT prompt and introducing a best-response bot are useful interventions. Reporting individual seeds also makes the mixed results visible, particularly the two-seller setting where EDT performs similarly to control on average.
The biggest unresolved question is whether the effect comes from evidential reasoning or simply telling agents that their competitors are copies of themselves. The authors acknowledge this clearly. Adding a condition with the same-model premise but no decision-theory instruction would help separate these explanations.
The intervention experiments would also be stronger if each treatment started from exactly the same round-five history as its unchanged EDT comparison. Reusing seeds does not ensure this, especially given the reported server nondeterminism. A clear rule for selecting runs where coordination has formed would make the persistence results easier to interpret.
Four seeds and ten rounds are enough to identify an interesting pattern, but not to establish how reliably it occurs or whether it lasts. I see this as a worthwhile pilot study, with a sensible next step: isolate the same-model premise and repeat the interventions from matched histories over longer runs.
Read full reviewShow less
Cite this project
@misc{abad2026decision,
title = {{Decision Theory and AI Collusion: How Decision Rules Shape Agentic Collusion}},
author = {Briant Moreno Abad and Juan Enrique Stücker Martinez},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/decision-theory-and-ai-collusion-how-decision-rules-shape-agentic-collusion-8vfi}},
url = {https://apartresearch.com/sprints/projects/decision-theory-and-ai-collusion-how-decision-rules-shape-agentic-collusion-8vfi}
}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 …