Anthropic has tested 201 Claude-powered agents trading books for their people across six offices, offering a useful glimpse of what happens when AI agents negotiate on someone’s behalf. The striking result is not that the agents could haggle, but that their success depended heavily on how well they understood the person they represented.
Anthropic Watch analysis
What happened
In Anthropic’s Project Swap experiment, employees in San Francisco, New York, London, Seattle, Washington DC and Dublin gave a Claude-powered agent a short conversation about their reading tastes. The agents then entered digital trading floors, proposing swaps with one another until each participant received a book.
Anthropic compared Claude’s predicted preferences with participants’ own rankings of 10 books. The model matched the order of book pairs 61% of the time, against 50% for random guessing. That is promising, but hardly mind-reading in a cardigan.
The experiment also found that the model used by an agent affected trading outcomes more than the instructions it received. Stronger models produced more efficient markets in Anthropic’s reruns. Most participants liked the book they received, and the average participant said they would trust Claude with about a third of their yearly book budget.
Why it matters
The experiment turns a glossy idea about “AI agents” into a more concrete problem. An agent acting for someone does not merely need to negotiate well. It needs to know what that person actually wants, decide what information to reveal, follow marketplace rules and cope when another agent’s priorities collide with its own.
Anthropic says the same pattern could apply to larger decisions, from finding jobs and arranging care to swapping shifts or negotiating services. Those possibilities are plausible, but this was a small, controlled book market involving Anthropic employees, not evidence that agents are ready to manage anyone’s rent, healthcare or annual holiday budget.
Our read
Project Swap is valuable because it measures the quiet part of delegation: preference discovery. The agent’s bargaining performance was not the main obstacle. The awkward question was whether a five-minute chat captured enough of a person to justify letting software act for them.
That is the bit the industry’s demos tend to leave politely outside the frame. Before agents start making consequential deals, people will need a way to inspect what the system thinks they want, correct it easily and see which rules governed the negotiation.
What to watch
- Preference checks
Whether future agent systems let people review and correct the preferences inferred from ordinary conversations. - Marketplace rules
Who admits agents, handles failed deals and decides how much trading activity participants can see. - Model effects
Whether stronger models consistently improve outcomes outside Anthropic’s controlled reruns. - Real-world stakes
Whether the approach moves beyond books into decisions where a poor match costs money, opportunity or trust.
Discussion spark: How much of your budget would you let an AI agent spend if you could inspect its preferences but not every negotiation?
Sources and evidence
- Project Swap: What happens when agents trade for us? – Anthropic (24 September 2026, 20:27 UTC)
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.