Anthropic runs agent-only book-swap market
In an internal experiment called Project Swap, Anthropic found that Claude agents trading books for 201 employees traded well, and that most lost value came from the agents not fully knowing what people wanted.
Anthropic has published results from Project Swap, a small barter market where Claude-powered agents traded books for their owners. The company describes it as a more controlled follow-up to Project Deal, its earlier experiment with agents dealing in a marketplace on people's behalf.
Each of the 201 participants, spread across six Anthropic offices, brought a book to give away. They had a short chat with Claude about what they like to read. Their agents then went onto a digital "trading floor" to pitch, haggle and swap with other people's agents until time ran out.
To measure how well the agents understood their owners, participants also ranked 10 books. After a roughly five-minute conversation, an agent's ranking matched its owner's on 61% of pairs. Anthropic called that "surprisingly good" for so short a chat.
According to Anthropic, the agents traded well once they were on the floor. The market fell short mainly because the agents lacked information about their participants, not because they negotiated badly.
The researchers then re-ran every trading floor dozens of times with different models and instructions. The underlying model affected negotiating outcomes more than the instructions did, and markets with stronger models were more efficient.
Most participants who read the book they received liked it. On average, participants said they would let Claude spend about a third of their yearly book budget.
Anthropic says the results point to two needs. Agent builders need ways to check that an agent understands its user. Market designers need rules on which agents can take part, what happens when a deal falls through, and how much market activity participants can see. The excerpt does not give full efficiency figures or name the models tested.