Anthropic tests AI agents that negotiate books

Anthropic put 201 employees and their AI agents in a book-trading market to see what happens when models make decisions in a marketplace. The result was clear: the agents could negotiate, but they did not always understand what each person wanted.

The experiment, called Project Swap, took place across six Anthropic offices: San Francisco, New York, London, Seattle, Washington D. C. and Dublin. Each participant brought a book they wanted to give away and had a brief conversation with Claude about their tastes and what they felt like reading.

The agent then entered a digital marketplace to propose trades, accept offers and negotiate with other agents. The goal was simple: each person should end up with a book they would enjoy more than the one they brought.

Claude understood people’s tastes fairly well, but not completely

To measure the quality of the decisions, participants ranked ten books according to their preferences. Claude never saw those lists. It only used the initial conversation to create an estimated ranking of all the available books.

Claude’s prediction matched the human choice in 61% of pairwise book comparisons. That is better than guessing at random, which would produce 50%, and it also beat simple methods based on popularity or on the books that similar readers tend to enjoy.

The conversation mattered. The median participant wrote 216 words across eight messages, and those who contributed around 300 words instead of 150 achieved approximately four percentage points more agreement between their tastes and the agent’s.

But a brief conversation does not capture every nuance. Someone may know they want to read something different without being able to explain exactly what “different” means. There were also cases where people received a book they had already read.

The main problem was not negotiation

Anthropic compared the actual result with the best possible allocation. If a system knew everyone’s preferences perfectly, market efficiency would reach 0.89 on a scale where 1 represents the ideal outcome.

The decentralized marketplace, in which the agents negotiated directly, reached 0.55 according to the participants’ actual preferences. To understand the cause of the gap, the researchers calculated what would have happened using Claude’s imperfect preferences, but with the best possible allocation.

That scenario reached 0.60. The conclusion matters: 85% of the distance from the optimal result was explained by a poor representation of people’s tastes, while only 15% was caused by how the market and negotiations worked.

In other words, the agent did not mainly fail because it could not bargain. It failed because it did not have enough information about the person it represented.

The model mattered more than the instructions

Anthropic reran the markets dozens of times using different models and instructions. Some agents were told to act “ruthlessly”, seeking only the best book for their user. Others were supposed to be “prosocial” and try to ensure that all participants ended up satisfied.

The difference between the two groups was small. Ruthless agents scored only 0.02 higher on Claude’s estimated preferences. By contrast, switching from a Haiku model to an Opus model improved the result by approximately 0.12 points.

In markets made up only of agents with neutral instructions, agents based on Haiku achieved an average efficiency of 0.75 according to their own estimated preferences. Those using Opus reached 0.88, close to the 0.95 optimum calculated using those same preferences. Sonnet fell between the two, while Fable came close to Opus, though slightly below it.

The agents also displayed recognizable behaviors: they often revealed which book was their favorite, compared offers, used time pressure and, in some cases, acted as intermediaries to help others complete a trade.

Would you trust an agent with part of your decisions?

Three weeks later, participants rated the book they had received at 7.2 out of 10. Nearly half said they liked it more than most of the books they choose for themselves.

When asked what share of their annual book budget they would leave under an agent’s control without the ability to review each purchase, the average response was approximately 30%. For comparison, they would hand around 40% to a friend who knew their tastes well.

Trust also depended on whether Claude had correctly summarized the initial conversation. Those who said the summary had not left anything out would delegate 34% of their budget. Those who spotted errors would delegate only 23%.

This points to a practical rule for any agent that will act on your behalf: before giving it autonomy, it should show you how it understood you and what decisions it would make with that information.

The study was small and controlled. All the agents were Claude models, behaved cooperatively and operated under rules set in advance. Participants were also Anthropic employees, so they probably trusted the technology more than the general public.

Even so, the experiment points to problems that will arise when agents negotiate jobs, purchases, shifts or services for us: how to verify that they faithfully represent their user, who is responsible when a transaction fails, what information should be visible and how to prevent thousands of agents from flooding a market with messages.

The idea of an agent negotiating on your behalf does not first run into a speed problem, but a representation problem. Before asking whether it can get you a better deal, you will need to know whether it really understands what kind of deal you want.