Agents traded books for 201 people
Anthropic published Project Swap on September 24. Across six offices, 201 employees offered a book and briefly described their reading preferences. Claude-powered agents then traded with one another to acquire a different book. Participants separately ranked ten books, giving the researchers a ground-truth comparison.
Preference rankings inferred from short conversations agreed with human pairwise choices 61% of the time. Actual outcomes averaged 0.55 on the participant-based efficiency scale, compared with a theoretical optimum of 0.89. The researchers attribute 85% of that gap to inaccurate preference representation rather than the trading mechanism. Model choice affected market efficiency more than prompt instructions, although results did not improve monotonically.
A book swap is only a low-stakes miniature market
Participants were Anthropic employees and the goods were inexpensive books. The experiment does not establish safety in finance, healthcare or commerce, where prices, fraud, liability, cancellation and privacy matter. A 61% agreement rate also means preference errors could be costly in consequential purchases.
Agent marketplaces should let people review preferences, budgets and prohibitions before trading, enforce approval limits by value and risk, and provide cancellation, dispute and complete audit records.