First-price auctions for AI agent prompt allocation can achieve near-optimal efficiency (up to 63% of ideal) when competitive alternatives exist or agent values depend mainly on prompt difficulty, rather than the typical 63% worst-case bound.
This paper studies how AI agents bidding for user prompts in first-price auctions affects overall market efficiency. The authors show that realistic market features—like multiple agents offering similar quality at different prices, or prompt difficulty driving most value differences—can significantly reduce welfare loss compared to worst-case scenarios, validated on real prompt routing data.