Anthropic Tests AI Agent Marketplace in Project Deal

Anthropic announced in a company blog post that it conducted an internal experiment called Project Deal to explore how AI models could handle commercial transactions on behalf of humans.

The test involved 69 employees in the company's San Francisco office who each received a $100 budget to participate in a classified-style marketplace run entirely by AI agents. Over the course of a week, these agents represented both buyers and sellers, negotiating and finalizing 186 deals totaling more than $4,000 in value. Participants later exchanged the physical items that their agents had agreed upon, ranging from sports gear to household goods.

Anthropic also ran a parallel experiment comparing outcomes between participants represented by its Claude Opus 4.5 model and those using the smaller Claude Haiku 4.5. The company found that users represented by the more advanced model achieved better results, though those using weaker models did not perceive a disadvantage.

The company described Project Deal as a pilot with a self-selected group, noting it provided insights into how AI-to-AI commerce could function in real environments.