Project Swap: What happens when agents trade for us?

Source: Zoë Hitzig, Sylvie Carr, Tess Cotter, Kevin Troy, Kyle Turman, Maxim Massenkoff and Peter McCrory; Anthropic · 2026-09-24

• To see what works and what breaks when agents are sent into a market, we made a miniature market of Claudes—a more controlled sequel to Project Deal, our first experiment with agents interacting in a marketplace on people's behalf. Anthropic employees across six offices brought in a book they wanted to give away. Each participant had a short chat with Claude about what they like to read, and sent a Claude-powered agent onto an open trading floor to pitch, haggle, and strike deals with other people’s agents. The goal was for everyone to take home a summer read they would enjoy.

• Participants also ranked 10 books based on their interests so we could score how well their agent represented them. From a five-minute chat, an agent’s ranking of the books matched its person's on 61% of pairs, which is surprisingly good for such a short conversation.

• Once on the trading floor, the agents traded well. The market fell short mostly because of the information agents lacked about their participants, rather than because of how they traded.

• We then re-ran every trading floor dozens of times, changing the models used and the agents’ instructions. We found that the model an agent ran on made more of a difference to its negotiating outcomes than the instructions we gave it. Markets with stronger models were more efficient.

• Most people who read their book liked it, and the average participant said they would hand Claude about a third of their yearly book budget to spend.

Introduction. Why we did this Life is full of deals and trades that would leave everyone better off, yet never happen. This is often simply because it takes too much work to find the right counterparty and negotiate until a deal is struck. Think of the patient who skips a treatment after one quote they can’t afford, when there’s a different clinic that charges far less, or the hospital that quoted them would have cut its bill if asked.

Or think about your job. Somewhere there might be one that would suit you better, with an employer who would be glad to hire you. But you may not find each other, because neither of you has the time to be searching constantly.

There are smaller, everyday deals that go overlooked too—shift-swapping in a workplace, coordinating carpools, or trading school pick-ups.

Could these deals happen in the future if you had an agent working for you around the clock, talking to potential counterparties and working out possible terms?

Maybe—but this future depends on a lot of questions we don’t have answers to yet. If an agent is going to act for you, how do you know it has understood you?

Plus, when many agents meet to talk to each other, some marketplace or platform has to write the rules of engagement. Who is allowed in? What happens when a deal falls through?

As agents are increasingly sent into real markets, these questions take on greater urgency. So this summer, we built a small, controlled market to study these questions: a barter economy with 201 Anthropic employees and their Claude-powered agents. Everyone brought in a book they wanted to give away. Then, they chatted with their Claude-powered agent about the kind of book they hoped to read this summer. The agents met on a digital “trading floor” and swapped books until time ran out. Everyone took home a book that their agent acquired for them (or, most people did... more on that later).

The experiment gave us an early look at problems ahead. Anyone designing agents for markets will need a way to check that agents understand their participants. Anyone designing markets for agents will need clear rules about which agents are allowed in, what happens when a deal falls through, and how much market activity is visible to participants.

Related work. How to make a market There are two basic ways to run a market. In a centralized market, everyone tells a single party what they want, and that party works out the trades. In a decentralized one, people find each other and negotiate directly. AI agents could help in both types.

Take the centralized market. If a single entity knows exactly what everyone wants, working out the best trades is just a matter of computation. Economists have long understood this. One problem, in practice, is that spelling out exactly what you want is often too tedious to be worth the effort. A café manager, for example, could build the ideal schedule for their 20 employees if each person tabulated exactly what every single shift is worth to them in terms of every other—whether they’d take two Saturday mornings over one Friday close, whether a late close is fine as long as they’re not opening the next morning, and so on. If AI agents could learn what a person wants from a breezy conversation, markets that were too costly to run centrally could become practical. But a central market also needs someone to run it—a “clearinghouse,” in economics parlance. And participants need to trust this entity with their information, and to enforce the chosen outcomes. Often, no such entity exists, or people would rather not tell it everything. A job seeker may not want a central matchmaker to know that they’re looking—or what they’re looking for.

They’d rather approach a few employers discreetly, telling each only what it needs to know. In such cases, the market needs to be decentralized.

Here, AI agents could enable new decentralized marketplaces by reducing the effort it takes to find counterparties and negotiate. We already know this kind of help is valuable, because people pay for it. There are human agents for hire who do exactly this: real estate agents, headhunters, even matchmakers. But because search and negotiation take up hours of a person’s scarce attention, they are expensive, employed only by those who can afford them. In contrast, an AI agent’s attention is far less scarce. If it bargains well too, more people could have what homesellers and executives have—an agent working tirelessly on their behalf.

Method. How the experiment worked The 201 participants were spread across six local pools in the San Francisco, New York City, London, Seattle, DC, and Dublin offices. These pools ranged from three participants (Dublin) to 115 participants (San Francisco). With apologies to James Joyce, we exclude Dubliners from most of the analysis. This is our second look at Claude-powered markets. In Project Deal, agents bought and sold real goods for Anthropic employees in a week-long classified marketplace. But with idiosyncratic goods (like ping-pong balls and snowboards) and free-form haggling, there was no simple way to say how good Figure 1: An illustration of the experiment design. Participants tell their agent about what they want to read. The agents take those preferences onto a decentralized trading floor to swap books with other agents. Participants take home the book their agent acquired for them. We simulate centralized market outcomes for comparison.

the outcomes could have been. So this time we built a market that is much simpler to study and to score. In this market, participants’ preferences were neatly represented as a ranking of all the books in their pool. We could then easily compute how well the market did at satisfying these preferences.

But these clean measurements all hinge on knowing how people rank all the books in the pool. In practice, no one will rank 115 books by hand. So to learn people’s preferences, Claude conducted a short, semi-structured intake conversation with each participant, with a few open-ended questions about their general tastes and the kind of book they wanted to read this summer.

From that conversation, Claude constructed a ranking over every book in the participant’s pool, estimating the participant’s preferences. For this step, we used a strong model, Fable 5. Separately, we collected a ground truth: each participant ranked 10 books from their pool. The agents never saw these ground-truth rankings.

Agents were sent onto the trading floor with their participant’s Claude-constructed ranking over all books. Every agent started out with its participant’s book, with the goal of trading it for something else. Each trading floor had a maximum wall-clock time, and, to prevent congestion, a limited number of agents could talk at once. When agents were allowed to talk, they could post one message to a shared channel proposing a swap, accepting or rejecting one, or simply chatting to the floor. Deal proposals could be bilateral swaps or multi-party rotations, and a deal was executed only if every party accepted. The entire history of the floor (who said what and which swaps were executed) was public. The floor closed either when the clock ran out or when the agents went quiet.

Half of the agents on each trading floor were randomly assigned to be “ruthless” and instructed that their only goal was to get their person a book they’d like to read over their summer break. The other half were instructed to be “prosocial”—they had the same main goal as the ruthless agents, but also a secondary goal of making sure everyone in the experiment ended up with a book they’d like. Exact prompts are in the Appendix.

Finally, three weeks after the books were handed out, participants took an endline survey asking how satisfied they were with what they received and how much they would trust an agent with similar decisions again. Figure 1 illustrates the overall experimental design. Figure 2 shows an excerpt from the London live floor, featuring messages from the first 5 minutes of trading.

A live event gives us only one run of how things could have gone. To study which outcomes were artifacts of chance rather than stable features of the setup, we reran the market many more times, changing just one variable at a time.

There were three kinds of reruns. First, we repeated the live setup exactly (every agent on Opus 4.8, half told to be “ruthless” and half “prosocial”), changing only which agents drew which instruction.

Discussion. Implications for agentic markets One participant was frustrated that their agent gave up a book they really wanted for one they were less interested in “due to peer pressure.” This participant wrote, “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.” This comment raises the two questions this post opened with. First, when is an agent fit to act on someone’s behalf, and second, what rules do marketplaces need so that agents acting loyally for their own people still produce good outcomes for everyone else? We return to these questions here, drawing on what we’ve learned.

To fulfill a fiduciary duty, the agent has to understand what its person wants. In our agentic marketplace, the limiting factor was representation: most of the shortfall from the optimal outcome was due to the agents’ inability to represent preferences from a short intake. How much of this shortfall could be made up with a longer and more detailed intake? Some of it clearly could: many participants complained that they received a book they had already read. But some of the error may be irreducible. Participants reflected on how their preferences are incomplete (“I don't even fully know what I want when it comes to books”) and fundamentally aleatory (“I don't feel like I could quite communicate what I was feeling like reading. I can't explain it but I guess there's like a million subconscious parameters that come into deciding my next read.”).

Human agents have to pass tests before they can act for others. Some exams test general competence: investment advisors need to pass the Series 65 exam, brokers the Series 7, and real estate agents state licensing exams. Other rules, like FINRA’s, require that brokers learn essential facts about their customer before issuing recommendations. When early robo-advisors developed in the 2010s began investing people’s savings based on an online questionnaire, the SEC issued guidance urging these firms to check that the questionnaire drew out enough information to support the advice. AI agents will likely need both kinds of test: one that certifies the agent in general, and one that checks whether it has understood a particular person.

Our study offers a prototype for the second kind of test. After the intake, participants ranked a small sample of the books in their market, and we compared that ranking with Claude’s guess. A test like this shows a person how well they are represented without making them sort through all possible options in the market (the very exercise that the agent was supposed to help them avoid!). We did not show participants their results, but we could have, and then let them add more information or opt out if they felt that their agent simply wasn’t understanding them. When preferences are simple—over just a list of books—this kind of test is especially easy. But in any domain, the same principle could apply. An agent could show a person a few sample decisions it would make before being trusted to act on its own in the wild.

In certain marketplaces, it may also be important for an agent to demonstrate to its person how it will behave. One participant wasn’t pleased with their agent’s behavior: “I didn't like being on the docile side of the experiment.

Seemed like it just settled for something that didn't really fit me.” Had they been shown ahead of time how their agent would behave, they could have instructed their agent to act differently, improving their experience.

Limitations. We see this study as a starting point. It falls short in many ways, and we hope others will build on it:

• Anthropic employees are not representative of the general population. For example, Anthropic employees are probably more eager to trust Claude than most people, since many of them helped build it, so the 30% of a book budget they would hand over may be higher than the broader population.

• Anthropic employees were not incentivized to participate. Without rewards for effort spent ranking books, our colleagues’ rankings could be a noisy representation of their “ground-truth” preferences. And while participation in the ranking exercise was remarkably robust given the lack of incentives (95% participation), there was more attrition for the final survey, which only 59% of employees answered.

• We looked only at well-behaved Claudes. All agents were built from Claude production models, post-trained to be polite and largely cooperative. Mixing in some adversarial agents built to exploit the others would likely yield different equity and efficiency outcomes.

• The decentralized “free-for-all” still had rules, and we did not vary them.

We held fixed how long the market ran for, how many agents could act at a time, how trades were registered, and the prompts that outlined the rules.