Anthropic has conducted one of the most realistic experiments in the so-called agent economy — an internal marketplace called Project Deal, where all deals were made exclusively by AI agents. This is not a simulation: agents bought and sold real things for real money, negotiating independently among themselves without human participation in the negotiations.
Anthropic has conducted one of the most realistic experiments in the so-called agent economy — an internal marketplace called Project Deal, where all deals were made exclusively by AI agents. This is not a simulation: agents bought and sold real things for real money, negotiating independently among themselves without human participation in the negotiations.
The experiment lasted about a week and took place within the company, Techcrunch writes .
69 employees received $100 each (in the form of gift cards), over 500 products were put up for sale, AI agents represented the interests of each participant, and as a result, 186 deals were concluded for over $4,000.
Before the start, agents even «interviewed» their owners to understand what they wanted to buy or sell, and on what terms. Then people were effectively removed from the process: the search for goods, negotiations, bargaining, and final deals took place autonomously.
This was not a «one-click marketplace.» Agents themselves found relevant offers, offered prices, received counteroffers, conducted full-fledged negotiations in natural language, and closed deals without pre-written scripts.
The items range from personal belongings to odd purchases. For example, one agent spent part of his budget on 19 ping-pong balls.
In other cases, prices for the same product could vary greatly depending on the terms of the negotiation—for example, the same item was sold at different prices in parallel scenarios.
The most important discovery of the experiment was the inequality between agents. Stronger models (such as Claude Opus) made more profitable deals, bought cheaper and sold more expensively, while weaker agents lost out—and their owners didn’t even notice.
In numbers, the difference could be a few dollars per transaction, but on a market scale, it creates a systemic advantage.
In fact, this is the first case where the quality of AI is directly converted into an economic result.
Another important point: the initial instructions for the agents had almost no effect on the result. Even if the user set an «aggressive» or «cautious» strategy: prices almost did not change, and the success of the deals did not either.
That is, prompt engineering does not mean market control, and the architecture of the model is more important than instructions.
Project Deal showed that AI can represent human interests, negotiate, optimize benefits, and operate under uncertainty. And all this without rigid rules or scripts. Anthropic directly notes that this format of deals «works» — agents were able to autonomously manage the market.
Project Deal is part of a larger strategy. The company has previously tested AI-stores without humans and experimented with agents that manage processes in the real world. And in parallel, Anthropic is developing the infrastructure for agents (Claude, agent workflows, etc.).
