Anthropic tests how Claude agents trade on behalf of people

Anthropic ran a controlled market experiment to see what happens when AI agents negotiate trades for people. In Project Swap, 201 Anthropic employees brought books they wanted to give away, described their reading tastes to Claude, and sent Claude-powered agents onto a digital trading floor to negotiate swaps with other agents.

The result is a useful early signal for agentic commerce. Anthropic says Claude inferred participants’ book preferences from short intake chats well enough to match their own rankings on 61% of book pairs, compared with 50% for random guessing. Once trading began, the agents generally negotiated effectively; the larger limit was how much they knew about the people they represented.

Anthropic also reran the markets many times with different models and instructions. Stronger models made the market more efficient, while changing agents from more self-interested to more prosocial instructions mattered less. Most participants liked the books they received, and the average participant said they would trust Claude with roughly a third of their yearly book budget.

The practical point is not books. It is that AI agents may soon bargain over shifts, services, purchases, jobs or other everyday exchanges. Anthropic’s experiment suggests the hard part will be proving that agents understand user preferences and setting clear rules for agent-run markets.