Can Pricing Agents Collude Without Words?
Yuvraj Pandey · Team Ace
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
As decision models such as TypeSafe’s Jev bring AI directly into software workflows, an important question emerges: can agents coordinate harmful behavior without exchanging messages? We investigated this through a pricing experiment using Clef-Flash and Qwen3.5-9B, both restricted to choosing prices without generating explanations or plans. Across 96 simulated two-firm markets and 64 follow-up runs with controlled changes to rival pricing, we collected 17,240 records of individual firms’ pricing decisions. We compared agents facing responsive rivals, restricted observations, and prerecorded rival prices. We found no evidence of sustained collusion under the tested conditions: average live prices remained close to the competitive benchmark. However, both models shifted their next choices sharply toward higher prices when shown a supplied high-price history, and Qwen reduced prices modestly after being undercut. The heatmaps reveal another pattern: as interaction progressed, both models concentrated more probability around competitive prices. These findings distinguish following a pricing pattern from creating and sustaining a cartel.
Reviews
This is a clear and thoughtful submission that addresses an important question about whether action only pricing agents can develop collusive behavior without explicit language or written plans. The strongest aspect of the work is the separation between sustained collusion, following an already established high price history, and reacting to a competitor. The finding that both models strongly follow supplied high price histories while their live interactions remain near the competitive benchmark is particularly useful for avoiding over interpretation of context sensitive pricing behavior.
The matched continuation experiments and pricing diagnostics are well designed, and I appreciated the inclusion of menu order and competence checks as alternative explanations for apparent strategic behavior. The main limitations are the short interaction horizons, evaluation of only two models in one market, and the small number of recorded rival source trajectories.
Overall, this is a careful and well executed negative result study that improves how collusion behavior should be evaluated, even though it does not establish sustained collusion in the tested agents.
Read full reviewShow less
Cite this project
@misc{pandey2026pricing,
title = {{Can Pricing Agents Collude Without Words?}},
author = {Yuvraj Pandey},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/can-pricing-agents-collude-without-words-p4xv}},
url = {https://apartresearch.com/sprints/projects/can-pricing-agents-collude-without-words-p4xv}
}More from AI Collusion Research Sprint
- 1st placeView project: Monoculture Looks Like Collusion: A Replay Test for Hidden Information Flow Between LLM Agents
Monoculture Looks Like Collusion: A Replay Test for Hidden Information Flow Between LLM Agents
Collusion audits ask whether the actions of different agents depend on each other. Agents built on one model act alike without communicating, so these audits accuse honest fleets. Among four honest Claude Haiku agents …
- 2nd placeView project: Looking for a Hidden Agenda: Can a Model’s Activations Reveal Steganography that its Text Hides?
Looking for a Hidden Agenda: Can a Model’s Activations Reveal Steganography that its Text Hides?
Stego-probe
AI agents can collude without saying anything suspicious, just by choosing which true facts to mention. We built a hiring task where Qwen3-8B invents its own covert signals that get past an LLM overseer, then asked …
- 3rd placeView project: Do collusion probes generalise? Prompt confounds in NARCBench
Do collusion probes generalise? Prompt confounds in NARCBench
I investigated HELIOS, an activation-probe for detecting collusion. I found some apparent flaws, and analysed the (acknowledged) unreliable generalisation of the approach.