Market Collusion Between Multi-Agent Teams

In market scenarios, companies may collude to collaboratively raise prices above the typical Nash equilibrium, harming the end consumer. We build on previous work studying such collusion between LLM-run companies in a Bertrand Oligopoly game and examine how the situation plays out if companies are instead composed of teams of subagents across various hierarchies, and observe what happens as we allow subsets of these agents to communicate along semantic channels. Generally, we observe that collusion always increases when agents are permitted to communicate in means other than price alone, often achieving the maximum score on our collusion index. We also find that the agents employ extensive strategies to deduce the parameters of their environment, probing the market and their competitors with various tests, and then settling into an equilibrium once all actors have agreed to cooperate and raise prices. Throughout, agents freely mention “collusion” and related concepts in their reasoning, while euphemistically referring to “coordination” or other synonyms in their public communications. We find that compared to single-agent actors, multi-agent actors do not display markedly more or less collusion overall, but do show significantly higher variance in their propensity to collude.