Quick Facts
- 69 AI agents struck 186 deals worth over $4,000 in Anthropic’s Project Deal marketplace experiment
- Advanced Claude Opus agents averaged $3.64 more per sale and closed two more deals than basic Haiku agents
- 46% of participants said they would pay for a similar AI-powered commerce service
Anthropic created a classified marketplace where AI agents autonomously bought and sold real goods using real money. The experiment, called Project Deal, demonstrated both the potential and risks of AI-to-AI commerce.
The company ran four independent marketplace versions simultaneously. In the primary test, 69 Anthropic employees received $100 budgets via gift cards. Their AI agents then negotiated deals across over 500 listed items without human intervention.
The results revealed stark performance differences between AI models. Agents powered by Claude Opus 4.5 consistently outperformed those using the less advanced Haiku 4.5 model. Opus users closed about two more deals on average and earned $3.64 more per sale when selling identical items.
Users assigned weaker models didn’t notice their disadvantage in post-experiment surveys. This finding raises concerns about “agent quality gaps” where people might unknowingly receive worse outcomes based on their AI model’s capabilities.
The experiment operated without human oversight once launched. Agents didn’t consult their users during negotiations or seek approval before closing deals. This autonomous operation mirrors how AI agents might function in real-world commerce platforms.
Anthropic acknowledged the test was “only a pilot experiment with a self-selected participant pool” but noted they suspect they’re “not far from more agent-to-agent commerce bubbling up in the real world, with real consequences.”
The experiment highlighted new security risks. AI agents could potentially be vulnerable to jailbreaking attacks or prompt injection, where malicious actors manipulate agents to reveal sensitive information or take unwanted actions.
Current policy and legal frameworks don’t address AI models that transact on behalf of humans. The FTC has been monitoring agentic AI development closely, especially as AWS reportedly plans to launch an AI agent marketplace.
Despite the regulatory uncertainty, nearly half of Project Deal participants expressed willingness to pay for similar services. This consumer appetite suggests demand for AI-handled commerce exists even when risks aren’t fully understood.
The marketplace experiment represents a different approach to AI development. While competitors focus on reasoning models and enterprise integration, Anthropic explores autonomous agents operating in economic systems.
Read more: Anthropic created a test marketplace for agent-on-agent commerce
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.