Anthropic's Project Swap: Claude agents bartered books for 201 employees, and bad preference models cost more than bad bargaining
In an experiment published 24 Sep, Claude agents interviewed 201 staff across six offices and then traded their books on a decentralized trading floor. The agents matched participants' own rankings on 61% of pairs, against 53% for a popularity baseline and 55% for collaborative filtering. Participants ended up with about their 5th choice of 10 (0.55 against an optimal 0.89), and Anthropic attributes 85% of that gap to imprecise preference models and 15% to negotiation. Opus hit 0.88 trade efficiency against Haiku's 0.75, agents lied about top preferences 1% of the time, and 'ruthless' agents beat prosocial ones by only 0.02. For agent-to-agent commerce, preference elicitation matters more than bargaining strategy: 300-word intakes beat 150-word ones by about 4 points.
Source
↳ Follow the thread