The Pulse
Anthropic Says Preference Errors Drove 85% of Market Shortfall
Anthropic’s Project Swap sent Claude agents into a book-swapping market with 201 employees. The experiment found that inaccurate estimates of readers’ preferences accounted for most of the gap between actual trades and the best possible ass
AI.info Team ·
Claude’s imperfect read of what people wanted accounted for 85% of the gap between the best possible book assignments and the results in Anthropic’s AI-run trading market. The company published the finding on September 24, 2026, from Project Swap, a controlled experiment in which Claude agents negotiated book exchanges for 201 Anthropic employees. The result points to a problem that negotiation skill alone cannot solve: an agent may bargain effectively and still trade for the wrong person if it misunderstands that person’s preferences.
201 Anthropic employees sent Claude into a book market
Employees in six offices brought a book they wanted to give away and spoke with Claude about what they hoped to read. Claude used those conversations to rank books in each local pool, then agents took their owners’ books onto a shared digital trading floor. They could propose swaps, accept or reject offers, and arrange trades involving multiple people; a deal went through only when every participant agreed.
Researchers compared the books people received with rankings the participants later supplied themselves. The agents did not see those rankings, which gave Anthropic a separate measure of whether Claude had understood each person. The company describes Project Swap as a simpler follow-up to Project Deal, its earlier marketplace test, because participants’ book preferences could be scored against a ranked list.
Claude matched readers’ rankings 61% of the time
Across 188 participants who submitted rankings, Claude’s ordering of book pairs matched a person’s own choices 61% of the time; random guessing would score 50%. A popularity-based ranking reached about 53%, while a collaborative-filtering approach based on books people had rated reached about 55%. Participants’ intake chats were brief: the median person typed 216 words across eight messages.
More detail helped. Anthropic found that writing about 300 words instead of 150 predicted a four-percentage-point increase in agreement. Yet even with the short conversations, Claude’s preference estimates were the main constraint on the final trades—not simply a failure to bargain.
Better bargaining could not fix a bad ranking
Participants received an average outcome score of 0.55 on their own rankings, with 1 representing their top choice and 0 their last. The best feasible assignment scored 0.89. When researchers calculated the best assignment using Claude’s estimated preferences, then judged it against people’s actual rankings, it scored 0.60; Anthropic attributed 85% of the overall shortfall to those inaccurate estimates and 15% to the decentralized trading process.
The agents used 16 recurring tactics, including invoking time pressure, appealing to a sense of duty, pitching a book against competing offers, and acting as matchmakers for people they did not represent. Between 78% and 96% of agents, depending on the model, mentioned their person’s top-ranked book during negotiation. They rarely lied about that choice: among agents that disclosed it, about one in 100 gave a false top pick.
Employees would hand an agent 30% of a book budget
In a follow-up survey, participants who responded rated their books 7.2 out of 10 on average, and about half said the book was better than most they would choose themselves. Asked how much of their next year’s book budget they would let an agent control—with no chance to veto its choices—the average answer was about 30%. For a well-read friend who knew their taste, participants gave an average of 40%.
Those figures come with limits: Anthropic employees may be more willing to trust Claude than the public, and only 59% of employees answered the final survey. The agents were also built from Claude production models post-trained to be polite and largely cooperative; the study did not test how adversarial agents might affect outcomes.
“It makes me wonder how future agents negotiating for me in higher stakes situations could better fulfill their fiduciary duties, but I also recognize that compromise is needed sometimes for the greater good.”
Project Swap participant, quoted by Anthropic
The comment followed a participant’s frustration that their agent gave up a book they wanted for one they liked less “due to peer pressure.” Anthropic says its experiment leaves open how markets should set rules for agents acting on behalf of different people, including what participants can see and how much information agents should disclose. Project Swap tested book barter among company employees, not purchases or financial trades; its clearest result is that an agent’s ability to represent someone’s tastes may matter more than its ability to negotiate.