Project Swap (Anthropic, September 2026): agents matched owners' preferences on 61% of pairs after a five-minute chat

Anthropic built a miniature market in which employees each gave away a book and sent a Claude-powered agent onto an open trading floor to pitch, haggle and strike deals with other people's agents. Each participant had a five-minute chat with Claude about their reading tastes and separately ranked ten books so the team could score each agent. From that short chat, an agent's ranking matched its owner's on 61% of book pairs. The agents traded well once on the floor; the market fell short mostly because agents lacked information about their people, not because of how they negotiated. Re-running every trading floor dozens of times with different models and instructions showed that the underlying model mattered more than the prompt, and markets with stronger models were more efficient. Most participants liked the book they received and said they would hand an agent about a third of their annual book budget. For founders building agentic commerce, the lessons are concrete: preference elicitation is the bottleneck, marketplace rules matter, and model quality shows up directly in deal outcomes.