Anthropic published Project Swap on September 24, 2026, a controlled experiment on how AI agents might trade on behalf of people. Two hundred and one employees across six offices each brought a book they wanted to give away, had a short conversation with Claude, and sent an agent representing them onto a trading floor.
Agents had limited time to pitch, negotiate, and propose exchanges. Deals could be bilateral swaps or multi-party rotations, but every involved agent had to accept before a trade executed. Participants also ranked ten books, giving the researchers a way to test whether an agent had captured its person’s preferences rather than merely completing a transaction.
Anthropic reports that the rankings inferred from a five-minute conversation matched participants’ own rankings in 61% of pairwise comparisons. That suggests a short intake can capture part of a person’s preferences, not that it creates a complete personal model. More expensive, sensitive, or irreversible decisions would require richer context and explicit confirmation.
The researchers reran many versions of the market while changing the model, instructions, and mixture of agents. The model an agent used generally affected negotiation outcomes more than instructions such as being “ruthless” or “prosocial,” and markets with stronger models were more efficient. Ruthless agents had a small advantage when optimizing only for their own participant, but the difference was modest.
The agents did more than match books. Researchers observed 16 recurring tactics, including invoking time pressure, positioning a book against competing offers, keeping waiting lists, and brokering multi-party trades. This means an agent market is not only a question of whether models can negotiate; it also needs rules about acceptable behavior, information visibility, and what happens when a deal fails.
A survey three weeks later found that most participants liked their books, and the average participant said they would hand Claude about a third of their annual book budget. That is a subjective response to a low-stakes experiment, not evidence that people would delegate financial, medical, or legal decisions in the same way. It is nevertheless a useful starting point for studying trust in delegated choices.
The practical lesson is that an agent acting for a person needs more than a prompt. The system must handle preference elicitation, authorization boundaries, market rules, and recourse after an action. Before an agent enters a real marketplace, users should be able to inspect the preferences it inferred, set acceptable trade boundaries, and retain confirmation and reversal options for consequential actions.
Project Swap remains a limited controlled experiment: the goods were books, participants came from one company, and the market rules were designed by the researchers. It helps surface new questions about agents in markets, but it does not show that agents are ready to make ordinary consumer or commercial decisions independently.
