Anthropic Created a Test Marketplace for Agent-on-Agent Commerce

Project Deal: An Experiment in Autonomous Commerce

Anthropic recently conducted an experiment called Project Deal, creating a classified marketplace where AI agents acted as both buyers and sellers. This pilot initiative involved 69 Anthropic employees, each provided with a $100 budget (distributed via gift cards) to purchase items from their colleagues. The experiment successfully facilitated 186 deals, totaling over $4,000 in value.

Key Findings and Observations

Anthropic ran four distinct marketplaces for this study:

• Real Marketplace: Utilized Anthropic's most advanced model, with deals being honored post-experiment.
• Study Marketplaces: Three additional marketplaces were used for research purposes.

Agent Quality and Outcomes

• Superior Outcomes with Advanced Models: When users were represented by more sophisticated AI models, they achieved objectively better results.
• Potential for 'Agent Quality' Gaps: A significant concern is that users on the losing end of transactions might not realize they are at a disadvantage due to disparities in agent capabilities.
• Instructional Impact: Initial instructions given to the agents did not appear to influence the likelihood of a sale or the negotiated prices.

Implications for Autonomous Agents

This experiment highlights the potential and challenges of agent-on-agent commerce. While advanced agents can drive better outcomes, ensuring transparency and fairness in these autonomous transactions will be crucial for widespread adoption. The discovery of potential 'agent quality' gaps underscores the need for robust evaluation and standardization in AI agent development.