Your Agent's First Guess Is You

The agents were better at bargaining than they were at knowing whom they represented.

That is the useful result inside Anthropic's Project Swap, a miniature book market run with 201 employees across six offices. Each participant brought a book to give away, spent about five minutes telling Claude what they liked to read, and sent a Claude-powered agent onto a trading floor. The agents pitched books, negotiated swaps, and sometimes arranged multi-party rotations. At the end, most people went home with a different book.

The market was designed so the researchers could separate two problems that usually arrive tangled together. First, had the agent formed an accurate picture of its person's taste? Second, given that picture, could it make good trades?

The first problem did most of the damage.

Participants independently ranked a sample of ten books. Claude's rankings agreed with theirs on 61% of book pairs, compared with 50% for chance. That is impressive for one short conversation and inadequate as a substitute for the person. When the researchers decomposed the gap between the best possible allocation and what the decentralized market achieved, imperfect preference representation accounted for 85% of the shortfall. Bargaining on the open trading floor accounted for the remaining 15%.

The agents had learned to trade a cleaner preference than the people actually had.

This moves the design problem earlier than the one in Shopping Agents Need A Receipt where we talked about how a receipt can preserve what happened after a delegated purchase: the request, constraints, influence, approval, and route back. Project Swap asks what happens before any of that, when a conversation becomes a model of the person and the model begins acting as if the conversion were complete.

Some errors were wonderfully ordinary. Participants received books they had already read. One disliked that their agent appeared to yield to peer pressure. Another reviewed the trading log, discovered that their agent had held a book for an hour before swapping it away, and bought that book separately afterward.

The replay surfaced a preference that the intake had never captured.

Preference is often described as if it were stored somewhere inside us, waiting for a sufficiently attentive intake form. Sometimes it is. I know which authors I follow, which subjects bore me, and whether I am willing to spend Saturday with 700 pages of maritime disaster.

Other wants form through contact. A cover catches the eye. A bookseller places two unlikely neighbors on a table. The back copy is terrible but the first page is alive. A friend says, with suspicious intensity, that the book is not usually your thing. Browsing can be where a preference happens.

One Project Swap participant named this directly. They trusted the agent to find and purchase a book, but still wanted exposure to reading culture: looking around, handling books, and reading their covers. In that case, the friction the market removed was part of what the person came to the market for.

The study should not carry more weight than it earned. Its participants were self-selected Anthropic employees trading low-stakes goods with agents built by one company. The agents were cooperative, the market rules were fixed, and only 59% of participants completed the final survey. A ranking of ten books is also not ground truth in any grand sense. It is another elicitation surface, completed at another moment.

Even within those limits, the experiment gives delegated interfaces a better first question. Before asking whether an agent is authorized to act, ask whether the person can inspect whom the agent thinks it is acting for.

There is evidence that the intake can improve. The preregistered experiments behind Generative Active Task Elicitation found that language models asking open-ended questions could elicit more informative preference specifications than user-written prompts or examples alone. Project Swap also found that longer intake conversations were associated with better ranking agreement.

More questions are not the same as complete understanding. A better interface would make the resulting model testable before it becomes powerful. Show a few sample decisions. Name the signals that mattered. Let the person say which inference feels wrong, which preference is firm, and which one is only a mood. Preserve a replay so an outcome can update the model instead of merely closing the transaction.

This has an older institutional echo. The U.S. Securities and Exchange Commission's 2017 guidance for robo-advisers warned that automated questionnaires needed to gather enough information to support suitable advice and address inconsistent responses. The technology was less conversational, but the boundary was already visible: a system can execute its method correctly while operating from a thin account of the client.

Agents will make this boundary feel friendlier. They can ask follow-up questions, remember earlier conversations, explain themselves in a familiar voice, and turn correction into dialogue. That may make them better representatives. It may also make a partial portrait feel finished sooner than a form ever could.

The practical move is small. Before I hand an agent part of my book budget, I would ask it to choose among three books, show me the reader it thought it was choosing for, and wait long enough for me to say: that person sounds plausible, but it is not quite me.