Anthropic built a marketplace and then handed the keys to Claude. No humans at the negotiating table — just AI agents buying, selling, and closing deals on their own. If that sounds like the setup to a science fiction short story, you’re not wrong to feel that way. But this actually happened, inside Anthropic’s San Francisco office, and the results are genuinely worth talking about.
What Was “Project Deal,” Exactly?
Anthropic ran an internal experiment they called Project Deal. The setup was straightforward: they created a real marketplace for employees in their San Francisco office, then tasked Claude — their own AI — with doing all the buying and selling. The twist was that Claude operated on both sides of the transactions. It wasn’t just a tool helping humans shop. It was the buyer and the seller, negotiating autonomously without human intervention at each step.
Sixty-nine employees participated. Each person was given a $100 budget. By the end of the experiment, the agents had completed 186 trades with a total value exceeding $4,000. That’s not a simulation. That’s real money moving around based on decisions made by AI agents.
Why This Is Different From a Chatbot Helping You Shop
Most people’s experience with AI in commerce looks like a chatbot suggesting products or auto-filling a checkout form. That’s AI as a helper. What Anthropic tested is something structurally different — AI as an independent economic actor.
When Claude was negotiating in Project Deal, it wasn’t waiting for a human to approve each move. It was reading the situation, deciding what something was worth, making offers, and closing transactions. Both the buyer-agent and the seller-agent were running on Claude. So in many trades, you essentially had Claude negotiating with itself, on behalf of two different human participants who weren’t in the room.
The goal, according to Anthropic, was to test economic theories about how AI agents interact when placed inside a real market structure. Not a theoretical model. An actual one, with actual stakes.
What the Numbers Tell Us
- 69 employees participated in the experiment
- Each started with a $100 budget
- 186 trades were completed
- Total trade value exceeded $4,000
Those numbers suggest the agents were active. With 69 participants and 186 trades, that’s roughly 2.7 trades per person on average. The agents weren’t sitting idle — they were finding deals, valuing goods or services, and executing transactions at a pace that suggests the system worked well enough to keep moving.
The Bigger Picture for Regular People
You might be wondering why any of this matters to you if you’re not an AI researcher or a tech investor. Fair question. Here’s how I’d frame it.
Right now, when you use an AI assistant to help you plan a trip or draft an email, you’re still the one making the final call. You approve the hotel booking. You hit send on the message. The AI is a very smart suggestion engine, but you’re the decision-maker.
What Anthropic tested is a world where that changes. Where your AI agent goes out, finds the best deal on something you need, negotiates the price, and completes the purchase — while you’re doing something else entirely. Agent-on-agent commerce means two AI systems working out the terms of a transaction without either human being present for the back-and-forth.
That’s a meaningful shift in how we might relate to money, purchasing, and even work. If agents can trade on your behalf, they can also make mistakes on your behalf, or be manipulated by a poorly designed agent on the other side. The efficiency gains are real. So are the new risks.
What Anthropic Was Really Testing
Beyond the practical mechanics, Anthropic was probing something more theoretical: do standard economic principles hold when the participants aren’t human? Do AI agents respond to price signals the way economic models predict? Do they find equilibrium prices? Do they behave strategically?
Project Deal was a small-scale attempt to get real data on those questions. Sixty-nine people and $4,000 in trades won’t settle decades of economic theory, but it’s a starting point. And starting points matter, especially when the technology is moving as fast as this one is.
For now, the experiment lives in a San Francisco office. But the questions it raises — about autonomy, trust, and what we’re comfortable letting AI decide for us — are ones the rest of us will be answering sooner than we think.
🕒 Published: