Anthropic's Project Deal: AI agents trading real goods in a controlled marketplace

Anthropic ran a closed experiment called Project Deal where AI agents represented buyers and sellers in a marketplace, completing 186 real trades worth over $4,000. The experiment was a pilot with 69 Anthropic employees, each given a budget, and the agents negotiated on their behalf. Anthropic described itself as "struck by how well Project Deal worked" — but the results also raised questions about how model quality affects real economic outcomes.

How Project Deal Worked

Anthropic ran four separate marketplace configurations with different model tiers. In the "real" condition, everyone was represented by the company's most advanced model. In other conditions, participants were represented by older or less capable models.

The result: when users were represented by more advanced models, they got "objectively better outcomes" — higher values for sellers, lower prices for buyers. But here's the striking part: the users didn't seem to notice when their agents were upgraded mid-experiment. They accepted the better outcomes without awareness of the quality difference.

This raises an uncomfortable question for AI developers: if the economic impact of better AI models is invisible to users, how will markets price AI capability? If buyers and sellers can't tell when they're getting a better deal, the market signals that usually drive competition may be dampened.

What This Means for AI Agents

Project Deal is a data point in the broader question of how AI agents will participate in economic activity. The experiment showed that AI-to-AI commerce is technically feasible today — agents can negotiate, agree on prices, and complete real transactions.

The more complex questions are about what happens at scale: How do agents represent user preferences and constraints? What happens when agents from different providers negotiate with each other? How do existing regulatory frameworks apply to AI-to-AI transactions?

Anthropic framed Project Deal as an internal research experiment, not a product. But the results inform how the company thinks about agentic commerce — a likely frontier for AI deployment in the coming years.