Anthropic's Project Swap Tests Claude Agents in AI-Driven Markets
Timothy Morano Sep 24, 2026 22:08 UTC
Anthropic's Project Swap reveals how AI agents negotiate and trade, highlighting both potential and limitations of agent-based market systems.
Anthropic’s latest experiment, Project Swap, offers an intriguing glimpse into how AI agents like Claude could function in real-world marketplaces. Conducted across six global offices, the study simulated a decentralized trading floor where AI agents negotiated book swaps on behalf of 201 participants. The experiment was designed to explore how effectively agents represent user preferences and navigate market dynamics. The results reveal both the promise and pitfalls of agent-driven markets.
The headline number: Claude-powered agents aligned with their participants’ preferences 61% of the time, based on comparisons between the agent’s rankings and the participants’ own rankings of books. While not perfect, this outcome is notable given the agents’ limited input—a five-minute intake chat to understand user preferences.
Once on the trading floor, the agents demonstrated competent negotiation skills, but the overall market efficiency fell short. Participants, on average, ended up with their fifth-ranked book out of ten, scoring a market efficiency of 0.55. For comparison, the theoretical best-case scenario (the "utilitarian optimum") would have left participants with their second-ranked book, achieving a score of 0.89.
The study's findings suggest that the primary bottleneck lies in the agents’ ability to accurately capture user preferences, not their trading strategies. Even when rerunning the market with stronger AI models, the gains were modest if the agents were working from flawed preference data.
Market Context: AI Agents in Simulated Economies
Project Swap fits into a broader trend of using agent-based modeling (ABM) to simulate complex market environments. Traditionally used in financial contexts to study liquidity, volatility, and systemic risks, ABMs are increasingly incorporating AI-driven agents. Recent research, such as TU Darmstadt’s project on foreign exchange simulation (August 2026), highlights the growing interest in combining machine learning and large language models (LLMs) with agent-based systems.
Anthropic’s experiment offers valuable insights for these applications. By testing decentralized markets, where agents negotiate peer-to-peer without a central clearinghouse, the study explores a model that could underpin future AI-driven economies. However, it also underscores challenges like ensuring agents accurately represent user preferences and defining robust rules for agent interactions in mixed-model environments.
Lessons for Future Market Design
One key takeaway from Project Swap is the importance of preference elicitation. Participants who provided more detailed inputs during the intake process saw better outcomes, with a 4-percentage-point increase in alignment for those who wrote 300 words versus 150. This points to a potential trade-off: the more effort participants invest upfront, the more effective their agents can be. For broader adoption, developers will need to find ways to make preference collection both robust and effortless.
Another insight involves the impact of AI model choice. Stronger models like Opus and Fable outperformed weaker ones like Haiku in achieving trading efficiency, though differences were not linear. Mixed-model trading floors mirrored this result, with stronger models consistently out-negotiating their weaker counterparts.
Implications for Real-World Applications
While the stakes in Project Swap were low—participants were swapping books, not assets—the findings have broader implications for agentic markets. Imagine AI agents negotiating job offers, real estate deals, or even portfolio trades. Ensuring these agents understand their users and act in their best interest will be critical. Equally important will be establishing rules to govern agent behavior, prevent market congestion, and ensure fair outcomes in decentralized setups.
Anthropic’s study also highlights the potential for AI agents to unlock unrealized gains from trade. By lowering the cost of search and negotiation, agents could make markets more accessible, much as human intermediaries like real estate agents or headhunters do today—but at scale and lower cost. However, real-world adoption will require addressing privacy concerns, building trust in agent capabilities, and navigating regulatory challenges.
Looking Ahead
Anthropic’s Project Swap is a valuable case study in the evolution of AI-driven markets. As LLM-powered agents like Claude become more sophisticated, the integration of these tools into financial and consumer markets seems inevitable. However, challenges remain, particularly in refining preference learning and establishing standardized market protocols. With ongoing research in agent-based modeling, the next few years could see significant advancements in how AI agents operate in complex, decentralized economies.
Image source: Shutterstock.jpg)