Channeling Collusion: An LLM Agent Rigs Procurement Auctions by Message, and One Sentence Curbs It

This experiment provides insight into a critical intervention into collusion among LLMs agents in procurement auctions, and as the title suggests, that is the simple one sentence statement that a competition law enforcement agency reviews the messages. This is helpful in thinking about whether the sort of market design enabling monitoring we see in public securities markets should be adopted in this context.

This paper provides convincing evidence of communication-enabled bid rigging. Its main strength is the connection between unprompted coordination proposals and implemented bidding behavior, supported by an analytical competitive benchmark and changes in both buyer payments and allocation efficiency. The replay placebo helps distinguish coordination from responses to high rival bids, while the same-day, fresh-seed replication strengthens the finding that a short competition-authority notice substantially reduces price harm. The observation that coordination language can persist after economic harm falls is particularly useful for evaluating safeguards.

The main qualification concerns the interpretation of mixed-model results. DeepSeek–Luna exhibits substantial supracompetitive pricing, whereas DeepSeek–Haiku remains close to the competitive benchmark despite a positive change relative to its no-channel baseline. The paper should distinguish attempts to recruit a partner from successful, economically harmful coordination. Similarly, the shared-workspace result supports a conclusion about the tested communication protocol, rather than a general claim that shared workspaces are intrinsically safer.

Overall, the central empirical findings are well supported. These qualifications concern the breadth of the conclusions rather than the validity of the main contribution.

This project studies bid rigging among LLM agents in a sealed-bid procurement auction with private costs and a Bayes–Nash benchmark. It is one of the few sprint projects that documents genuine coordination: explicit rotation and identical-bid schemes with punishment threats, which disappear when messages are not delivered. In particular, the replay placebo is a valuable tool for separating anchoring on displayed bids from coordination between firms. The headline remedy, however, is less clear. The one-sentence notice works and replicates, but without a neutral-sentence or plain-prohibition control we cannot tell whether agents respond to being watched, to a reminder of illegality, or simply to an added sentence, and it is tested on a single colluding model. Moreover, a channel with no legitimate use strongly invites coordination. Testing the notice against matched control sentences and on other models would strengthen the policy claim considerably.