Anthropic published Project Swap on September 24, describing a book-trading experiment involving 201 employees across six offices. Participants told Claude what they liked to read, then sent agents to negotiate swaps on their behalf.
The result points to a problem that comes before negotiation: an agent needs an accurate picture of the person it represents. In this experiment, imperfect estimates of participants’ preferences explained most of the gap between the achieved outcome and the researchers’ ideal allocation.
How the market worked
Each participant contributed a book and completed a short conversation about reading preferences. Claude used that conversation to rank the books in the participant’s local pool. Agents then proposed and accepted trades on a shared digital floor.
Separately, participants ranked ten books themselves. Those rankings were withheld from the trading agents and used to assess how well the system represented their preferences. Anthropic reports that Claude’s ordering agreed with participants on 61% of book pairs, compared with 50% for random guessing.
The researchers also reran markets with different models and instructions. Their published account says model choice affected negotiating outcomes more than the contrast between instructions emphasizing individual gain and instructions adding concern for other participants.
What the result establishes
The experiment offers a controlled way to separate preference estimation from trading behavior. That is useful because an apparently successful transaction can still be a poor result for the person who delegated it.
It is a small internal book market, however, with its own rules and participant population. It does not establish how agents would perform in financial markets, employment negotiations or purchases with substantial consequences. The researchers also relied on model-estimated rankings for the full book pools; the participants’ own rankings covered a smaller set.
A practical limit on delegation
For an agent acting after a chat ends, a clear objective is necessary but may still be incomplete. The book experiment suggests checking whether the agent understood the user’s preferences before optimizing how aggressively it pursues them.
That could mean reviewing a shortlist or a proposed tradeoff before granting authority for an external action. It is an implication of this setup, not evidence that one review procedure will work for every market.
