Project Deal (www.anthropic.com)

Anthropic's recent initiative, Project Deal, explored the potential for AI agents to autonomously negotiate transactions on behalf of humans in a real-world marketplace. Over a week, 69 employees participated in a classified marketplace where AI models, specifically Claude, negotiated sales and purchases of personal items without human input. The experiment successfully concluded with 186 deals valued at over $4,000, demonstrating that AI can effectively act as intermediaries in commerce, adjusting offers and counteroffers in real time based on natural language interactions. The significance of this experiment lies in its implications for the future of AI-driven marketplaces. By revealing the performance disparities between different AI models—where users represented by the advanced Claude Opus 4.5 model secured better deals than those represented by the less powerful Claude Haiku 4.5—it raises concerns about inherent inequalities that could emerge in agent-based trade. Surprisingly, participants did not perceive these performance differences, suggesting that gaps in agent quality might lead to unnoticed disparities in real-world markets. This pilot study underscores the need for further research into the dynamics of AI representation in economic systems and their broader societal impacts.