Project Swap: What happens when agents trade for us? (www.anthropic.com)

Anthropic has launched Project Swap, a controlled experiment that explores how AI agents, powered by Claude, can negotiate and trade on behalf of users in a marketplace setting. In this mini-market, 201 Anthropic employees exchanged books after a brief conversation with Claude to express their reading preferences. Remarkably, agents matched participants' book rankings with 61% accuracy based on those short exchanges, showcasing the potential for AI to understand nuanced human preferences quickly. The experiment highlighted both the strengths and challenges of using AI in real markets. While agents effectively facilitated trades, the limitation was identified in the agents' understanding of their participants, raising questions about what information is essential for effective negotiation. The results suggest that enhancing models could significantly improve trading efficiency, as stronger models outperformed weaker ones in terms of negotiating outcomes. Such advancements may pave the way for AI agents to tackle everyday transactions, whether in healthcare, job markets, or simple exchanges, potentially streamlining processes that currently require significant human effort and time. This research emphasizes the urgent need to develop trust and clear operational rules for AI-run marketplaces as their use gains traction in the real world.