A market made of books and agents

Anthropic asked 201 employees across six offices to bring in a book they were willing to trade. Each person had a short conversation with Claude about what they liked to read. A Claude-powered agent then entered a shared digital trading floor, pitching books, accepting swaps and arranging multi-party deals on that person's behalf.

The setup was deliberately small and measurable. Participants separately ranked a sample of ten books, giving the researchers a human reference point that the agents never saw. Anthropic then replayed the markets many times with different models and instructions. The question was not simply whether agents could reach deals. It was whether those deals reflected what people actually wanted.

Claude had a partial map of each reader

From a median intake of 216 words across eight messages, Claude's ordering agreed with a participant's ordering on 61% of the book pairs they ranked. Random ordering would score 50%. A popularity baseline reached about 53%, and a collaborative-filtering method reached about 55%. The conversation extracted real information, but it still left substantial uncertainty.

That missing information shaped the result. On a scale where 1 means receiving a top-ranked book and 0 the last-ranked one, the best possible assignment scored 0.89. The decentralized agent market averaged 0.55 on participants' own rankings. A hypothetical best assignment based on Claude's inferred rankings reached only 0.60. Anthropic calculates that 85% of the gap from the optimum came from the preference model, while the bargaining process accounted for the remaining 15%.

Put plainly: the agents were fairly good at trading the preferences they had. The larger problem was that those preferences were only an approximation. Better negotiation cannot recover information that never entered the system.

Model choice mattered more than the personality prompt

In 80 neutral-instruction reruns, markets using stronger models performed better when outcomes were scored against Claude's own rankings. All-Haiku floors averaged 0.75, compared with 0.88 for all-Opus floors and a computed optimum of 0.95. Mixed-model floors did not erase the advantage: Opus agents generally finished ahead of their weaker-model counterparts.

Instructions had a smaller effect. Agents told to be ruthless scored about 0.02 higher on Claude's rankings than agents also asked to consider everyone's outcome. Prosocial agents were twice as likely to accept a book lower on their own list as a sacrifice, though both types of sacrifice were rare. By comparison, moving from Haiku to Opus shifted outcomes by 0.12 on the same internal scale.

Those differences largely faded when researchers evaluated results against people's own rankings. The paper says the study was not well powered to detect small design effects through the noise of imperfect preference estimates. That is an important limit, not evidence that agent design never matters.

The useful lesson is a test before delegation

Most follow-up respondents liked their book, giving it an average score of 7.2 out of 10. They said they would let an agent control about 30% of their annual book budget, compared with about 40% for a well-read friend. Only 59% answered that final survey, and all participants worked at the company that built Claude. The figures are interesting, not representative measures of public trust.

Anthropic's broader proposal is more useful than the trust number. Before an agent acts, it could show a person a small set of sample decisions. Disagreement would reveal that the intake is incomplete, giving the person a chance to add context or opt out. Logs could then make the agent's choices reviewable after the fact.

Books kept the stakes low. Real markets add money, identity, adversarial agents, delivery failures and legal duties. Project Swap does not show that autonomous agents are ready to negotiate a job offer or medical bill. It shows a narrower design problem: an agent can bargain competently and still pursue the wrong version of its owner's wishes.

Sources

  1. Anthropic Research: Project SwapPrimary September 24 report describing the live book market, reruns, headline results, limitations and implications proposed by Anthropic's economics team.
  2. Anthropic: Project Swap report and appendixPrimary 26-page paper with experiment design, prompts, scoring choices, statistical results, survey text and limitations.
  3. Anthropic: Project DealPrimary account of Anthropic's earlier agent marketplace experiment, used only to establish how Project Swap differs from the prior work.