Anthropic tests AI agents in a trading market

Anthropic published Project Swap, an experiment in which 201 employees used Claude-powered agents to trade books with one another. Participants briefly described their reading preferences before their agents entered a digital market to negotiate swaps.

From the short intake conversations, the agents' rankings matched participants' own rankings on 61 percent of book pairs. Anthropic then reran the markets with different models and instructions, finding that model capability affected negotiating outcomes more than the difference between its tested instruction styles.

The experiment used books and a controlled workplace setting rather than financial assets. Its useful result is narrower: an agent can negotiate effectively while still representing an incomplete model of what its user wants, making preference elicitation and market rules separate design problems from bargaining skill.