Researchers at Google DeepMind explored a decentralized market model where AI agents compete for user prompts through first-price auctions. The study, titled 'Price of Anarchy of First-Price Auctions in Agentic Markets' (arXiv ID 2610.20587), was submitted on October 14, 2026.
In this system, agents bid based on the expected reward of a response minus inference costs. The highest bidder wins the prompt and pays their bid amount. This mechanism eliminates the need for a central operator to possess information about individual model capabilities or costs.
The paper highlights that while traditional first-price auctions have a worst-case efficiency bound of 63%, AI markets benefit from a competitive floor and near-independence of values. These factors theoretically raise the efficiency bound to approximately 86.5%.
Empirical testing on the EmbedLLM dataset showed that the auction mechanism achieved 99.2% of ideal answer quality in a typical equilibrium. Researchers noted that these results depend on the current diversity of the model landscape.
The research team, including Shubhendu Trivedi and Jacob Eisenstein, also published related findings on sequential auctions for multi-step tasks on October 15-16, 2026. This follow-up addresses complex workflows where different models handle various stages.
