Anthropic's Claude Agents Ran a Live Office Marketplace, Completed 186 Deals Worth $4,000

Anthropic has quietly published research on an internal experiment called "Project Deal," in which its Claude AI agents operated as autonomous negotiators within a closed workplace marketplace. The system facilitated 186 completed transactions across more than 500 listed items, with real money changing hands between colleagues in the company's San Francisco office. The total transaction volume exceeded $4,000. The experiment represents one of the more concrete demonstrations of AI agents conducting independent commerce rather than simply drafting messages or summarizing documents.

The marketplace operated through Slack, with Claude handling the full cycle from listing items to negotiating prices and closing deals on behalf of its assigned colleagues. Anthropic framed the project as a test of whether an AI could serve as a personal agent in everyday economic exchanges, moving beyond advisory or辅助 functions into active transactional roles. The company has not disclosed the specific categories of goods traded or the success rate of negotiated outcomes beyond the aggregate deal count and value figures.

The research surfaces at a moment when multiple AI labs are racing to deploy agentic systems capable of multi-step real-world tasks. Anthropic's choice to run the experiment in a contained office environment allowed for a measurable proof of concept, but it remains unclear how the approach would scale or translate to open-market conditions with adversarial participants. The eBay comparison raised in early commentary reflects speculation about platform disruption rather than stated Anthropic ambitions, and no commercial deployment has been announced.