Claude agents match 61% of book preferences

Anthropic’s Project Swap sent Claude-powered agents representing 201 employees into a book-trading market, where five-minute intake chats matched participants’ rankings on 61% of book pairs. The official experiment report says stronger model choice mattered more to negotiating outcomes than instructions, while participants said they would hand an agent about 30% of their yearly book budget. Builders designing agents that act for users should measure preference capture before polishing negotiation: the archive’s Business Arena reliability test likewise treats long-horizon agency as a capital-preservation problem, not a chat-quality demo.