Anthropic’s AI-Powered Classified Marketplace: Real Deals, Real Money, No Humans Needed

Anthropic has conducted an experiment demonstrating that artificial intelligence agents can successfully negotiate and complete real commercial transactions on behalf of humans. In a controlled marketplace setting, AI agents representing both buyers and sellers struck 186 deals involving physical goods and actual monetary value, totaling over $4,000 in transactions.

The initiative, dubbed Project Deal, involved 69 Anthropic employees from the company’s San Francisco office who agreed to participate in a classified marketplace modeled after platforms like Craigslist. Each participant’s AI agent was given a budget of $100, distributed via gift cards, to spend on items listed by their coworkers. Prior to the experiment, Claude interviewed employees to determine which personal belongings they might be willing to sell and what types of items they would consider purchasing.

Anthropic ran four separate versions of the marketplace to study different variables. One version used the company’s most advanced AI model to represent all participants, with deals being honored in real money after the experiment concluded. The other three versions were designed for comparative analysis, allowing researchers to study the impact of model capability on transaction outcomes.

According to Anthropic, participants represented by more advanced AI models achieved “objectively better outcomes” in the negotiations. However, users did not appear to perceive differences in performance between agents, suggesting a potential lack of awareness when they were at a disadvantage in agent-mediated exchanges. This dynamic raises concerns about what the company describes as “‘agent quality’ gaps,” where individuals represented by less capable models might not realize they are receiving suboptimal results.

The initial instructions provided to the AI agents did not significantly influence the likelihood of a sale occurring or the final prices negotiated. Anthropic noted that the experiment was limited to a self-selected group of employees and characterized it as a pilot study rather than a broadly representative market test.

Despite its limited scope, the company expressed surprise at how effectively the agent-to-agent commerce system functioned. The successful completion of real deals for tangible goods using actual funds suggests that AI-mediated transactions could become a viable component of future economic activity, particularly as agent capabilities continue to advance.