Project Swap: What happens when agents trade for us?
Anthropic conducted "Project Swap," an experiment where Claude-powered AI agents traded books on behalf of 201 employees to study agent behavior in miniature markets.
Intelligence analysis by Gemini 2.5 Flash
The research found that AI agents effectively represented user preferences and traded well, with the strength of the underlying AI model being more critical to negotiation outcomes than specific instructions. The primary limitations stemmed from agents lacking comprehensive information about their participants, rather than their trading capabilities.
Imagine you have a smart robot friend who loves books. You tell it what kind of books you like in a quick chat. Then, this robot goes to a special online party with other robot friends, each trying to swap books for their human. Your robot tries to get you a book you'll love, and it's pretty good at it! The smarter the robot, the better it trades, but sometimes it just needs more clues from you about what you really want.
Analysis
Project Swap
Anthropic's "Project Swap" was a controlled experiment designed to explore the dynamics of AI agents operating in a market environment. Building on a previous initiative, Project Deal, this iteration involved 201 Anthropic employees across six offices. Each participant brought a book they wished to give away and engaged in a brief conversation with a Claude-powered agent about their reading preferences. These agents were then deployed onto a digital trading floor to autonomously pitch, haggle, and strike deals with other agents, aiming to secure a desirable summer read for their human counterpart. The experiment provided a valuable, early look into the complexities and potential pitfalls of agent-driven markets.
Claude's Performance
The Claude agents demonstrated a surprising ability to represent their participants' interests, with their book rankings matching human preferences 61% of the time after just a five-minute chat. Once on the trading floor, the agents proved to be effective traders. A key finding was that the underlying model's strength had a greater impact on negotiation outcomes than the specific instructions given to the agents, indicating that more powerful AI models lead to more efficient markets. The primary shortcomings of the market were attributed to the agents' limited information about their participants, rather than their trading proficiency, highlighting the importance of robust preference elicitation.
Decentralized Markets
The motivation behind Project Swap stems from the observation that many mutually beneficial deals and trades never materialize due to the significant effort required to find counterparties and negotiate terms. Examples range from patients finding affordable treatments to individuals discovering better job opportunities or simply coordinating everyday tasks like carpools. The research suggests that AI agents, with their less scarce attention compared to human agents, could significantly reduce these search and negotiation costs. This capability could enable the creation of new decentralized marketplaces where agents facilitate transactions, overcoming the practical challenges of centralized systems that require participants to trust a single entity with all their information.
Key points
- Anthropic conducted "Project Swap," an experiment where Claude AI agents traded books for employees.
- Agents successfully represented participant preferences (61% match) and traded effectively on a digital floor.
- The strength of the AI model was more influential on negotiation outcomes than specific instructions.
- Market inefficiencies primarily arose from agents lacking complete information about their human participants.
- AI agents could potentially enable new decentralized markets by reducing the effort required for search and negotiation.
If AI agents can effectively understand human preferences and negotiate on our behalf, they could unlock countless beneficial deals and opportunities that are currently too time-consuming or complex for individuals to pursue. This could lead to more efficient markets, better access to services, and improved personal outcomes across various aspects of life.
The experiment highlighted challenges, particularly ensuring agents fully understand their participants' nuanced desires and establishing clear rules for agent interactions in a marketplace. Without robust mechanisms for preference alignment and market governance, agents could make suboptimal decisions or create unforeseen complications in real-world transactions.
