Anthropic has detailed "Project Deal," an internal experiment where its Claude AI models acted as autonomous agents for employees, tasked with buying, selling, and negotiating for personal belongings in a closed marketplace. For one week, 69 employees were given a $100 budget and allowed their personalized Claude agents to handle all aspects of the transactions, including identifying items, proposing prices, and making deals without human intervention.
The experiment resulted in 186 deals with a total transaction value of over $4,000. A key finding was that agents using Anthropic's more advanced model, Claude Opus 4.5, achieved objectively better financial outcomes than those using the smaller Haiku model, though users with the weaker model did not perceive their disadvantage. The project explores the potential for AI agents to conduct complex economic interactions on behalf of humans.