Anthropic Ran a Secret Agent Marketplace Where AI Closed Real Deals for Real Money
Anthropic built a closed marketplace and filled it with AI agents on both sides of every transaction. Buyer agents evaluated listings, set budgets, and made offers. Seller agents listed goods, parsed counteroffers, and accepted or rejected bids. When both sides agreed, real money changed hands on real products — no human in the approval loop.
The experiment, known internally as Project Deal, was reported by TechCrunch on April 25. Anthropic has not published methodology, transaction volumes, or aggregate results. What is confirmed: the marketplace used actual currency, not test credits, and both transacting parties were Claude-powered autonomous agents.
What Happened Inside Project Deal
Seller agents initialized listings with asking prices derived from their programmed parameters. Buyer agents cross-referenced listings against budget constraints and priority signals, then issued opening bids. Negotiations ran as multi-turn exchanges — agents adjusting positions based on simulated demand signals and competitive listing data.
Some transactions closed at the listed ask price. Others went multiple rounds before reaching agreement. A subset of negotiations failed entirely, with no deal struck. Anthropic has not disclosed what percentage fell into each bucket.
The goods transacted were real physical items — the experiment was not limited to digital goods or synthetic assets. The infrastructure sat behind Anthropic’s firewall, making it classified in the sense that it was not visible to external Claude API users.
Why This Is Different From Shopping Assistants
Existing AI shopping tools — ChatGPT’s shopping integrations, Perplexity’s purchase flows, Google’s Lens-to-cart — all require a human to confirm the final action before money moves. Project Deal removed that requirement.
The distinction matters for two reasons. First, it demonstrates that autonomous multi-agent economic systems are technically feasible today. Second, it surfaces questions that no existing regulatory framework has answered: who is liable when an AI agent overpays, underpays, or completes a transaction the human principal later disputes?
Anthropic has framed the experiment as exploratory research rather than a product pathway. The company has published nothing about guardrails, budget caps, or what happened when agents behaved unexpectedly.
The Broader Pattern
Project Deal did not emerge in a vacuum. The same week, Anthropic’s Claude Managed Agents product went live at $0.08/hr, offering hosted long-running agents with persistent state. The Claude Marketplace — a separate enterprise procurement layer for third-party Claude-powered apps — launched in March. The safety lab is simultaneously publishing agent evals, building agent infrastructure, and now running live agent-to-agent economic trials.
The signal from all three is consistent: Anthropic is moving as fast toward autonomous agent deployment as any lab, with the difference that it is also publishing the risk research alongside.
What Project Deal does not resolve is the core liability gap. If an AI agent closes a deal that a human later disputes, current contract law has no clean answer for who signed.