Key Highlights:

• Anthropic ran a pilot marketplace where AI agents bought and sold real items for users.
• The experiment completed 186 deals worth more than $4,000.
• Advanced models delivered measurably better negotiation outcomes.
• Users often did not notice differences in agent performance.

Anthropic tested whether AI agents could negotiate and complete real purchases on behalf of people. In its internal pilot called Project Deal, agents represented both buyers and sellers and successfully executed 186 real transactions worth over $4,000.

The experiment involved 69 employees, each given a $100 gift-card budget. Their AI agents handled the negotiations in a controlled classified marketplace environment.

What is Anthropic’s Project Deal marketplace experiment?

Anthropic created a small digital marketplace where employees interacted through AI agents instead of negotiating directly with each other. These agents acted as intermediaries. They searched listings, negotiated prices, and finalized agreements.

Importantly, one version of the marketplace operated under real conditions. Deals completed there were honored after the test ended. Meanwhile, three additional versions ran as comparison environments for research.

The company described the setup as a limited pilot with a self-selected participant pool. Still, results showed strong evidence that agents can coordinate real transactions independently.

Did stronger Anthropic models deliver better outcomes?

Anthropic reported that users represented by more advanced AI models achieved objectively better deal outcomes. However, participants rarely noticed these differences during the experiment.

This finding introduces what researchers called “agent quality gaps.” In such scenarios, users may receive weaker negotiation results without realizing their agent performed worse than others.

The outcome suggests future digital marketplaces could include hidden performance disparities between AI assistants acting on behalf of users.

Did agent instructions change negotiation behavior?

Researchers also tested whether different starting instructions influenced how agents negotiated. Surprisingly, the initial prompts did not significantly affect sale likelihood or final prices.

This result indicates that agent capability mattered more than instruction style in determining transaction success inside the marketplace.

Why Anthropic’s agent marketplace experiment matters

Project Deal offers an early glimpse into how autonomous agents could reshape online commerce. Instead of browsing and negotiating manually, users may soon rely on AI representatives to handle transactions automatically.

Anthropic’s pilot shows that agent-to-agent commerce is already technically possible in controlled settings. As the company continues exploring these systems, Anthropic’s findings highlight both opportunity and risk in future AI-mediated marketplaces.