Anthropic created a test marketplace for agent-on-agent commerce

Project Deal: Anthropic Tests an AI Agent Marketplace

AI research firm Anthropic has revealed the results of an internal experiment called Project Deal, which created a classified marketplace where AI agents autonomously negotiated and executed transactions on behalf of human users. The test, which involved real goods and money, serves as a proof-of-concept for agent-on-agent commerce, demonstrating the technical feasibility of such systems while also surfacing critical questions about market fairness and transparency in an AI-mediated economy.

Experiment Details and Key Metrics

While described as a pilot experiment with a limited participant pool, the project generated significant activity and data. Anthropic set up four distinct marketplaces to study agent behavior under different conditions, with one 'real' marketplace using its most advanced model to execute deals that were honored post-experiment. The core findings from the project underscore its operational success.

• Participants: 69 Anthropic employees
• Budget: $100 per participant (via gift cards)
• Total Deals: 186 successfully negotiated transactions
• Total Value: Over $4,000 in goods exchanged
• Key Variable: The AI model version representing the user

The 'Agent Quality Gap' and Market Implications

The most notable outcome from Project Deal was the identification of an "agent quality gap." The experiment confirmed that users represented by more advanced AI models achieved objectively better outcomes in their negotiations. Critically, however, users on the losing end of these deals often did not realize they had received a worse bargain, raising significant concerns about information asymmetry and fairness. This suggests that as agent-based commerce becomes more widespread, access to cutting-edge models could become a decisive economic advantage, creating a new form of digital inequality that is largely invisible to those it affects.

While proving the viability of agent-to-agent commerce, Anthropic's experiment also provides an early warning about the potential for new forms of digital inequality, where the quality of one's AI agent—not just their negotiating skill—determines economic outcomes.