Anthropic tests a marketplace where AI agents haggle and close deals on their own

Anthropic has shown that a market where AI agents trade with each other, without a human involved at every step, no longer looks like science fiction. In Project Deal, the company built an internal classifieds marketplace where Claude represented both sellers and buyers: it posted listings, negotiated, made counteroffers and recorded deals. Importantly, this was not a simulation on test data: employees really did exchange goods and money once the experiment ended.

The experiment ran among 69 Anthropic employees at the company's San Francisco office. Each received a notional budget of $100, which was later settled with gift cards based on their purchases and sales. Before launch, Claude conducted a short interview: what the person was willing to sell, what they wanted to buy, at what price, and how exactly the agent should negotiate. Each participant then got a separate agent with its own system prompt. Bidding took place in Slack channels, and humans did not confirm each deal by hand: once the run started, the agents acted on their own.

Anthropic ran four versions of the market in parallel. One counted as "real": participants later exchanged goods based on its results. The other three served as comparisons. In two runs, Claude Opus 4.5 represented everyone; in the other two, participants were randomly split between Opus 4.5 and the lighter Haiku 4.5 model.

In the "real" run, the agents closed 186 deals across more than 500 listings, worth just over $4 thousand. These were not instant one-click purchases: the agents had to find matching interests, argue over price, answer counteroffers and bring each conversation to an agreement.

The main takeaway was not that the AI managed at all, but that model quality directly affected the outcome. According to Anthropic, users with Opus closed roughly two more deals on average than users with Haiku. When Opus sold an item, the price was on average $3.64 higher. Overall, Opus brought in about $2.68 extra as a seller and, as a buyer, cut the price by roughly $2.45. Individual cases were starker: the same lab-grown ruby sold for $65 through an Opus agent but only $35 through Haiku, and a broken folding bike went for $65 in one run versus $38 in another.

Curiously, the style of instructions barely changed the outcome. Some employees asked their agents to be gentle and friendly; others told them to bargain hard and open with lowball offers. This had no statistically significant effect: aggressive prompts neither raised the chance of a sale nor consistently won a better price.

Other quirks of agent-driven trade did show up. One agent bought a person a snowboard nearly identical to the one they already owned, apparently taking the owner's tastes far too literally. Another, at an employee's request, picked a gift "for Claude itself": a pack of 19 ping-pong balls for $3. Some deals were about experiences rather than things: for example, agents arranged a free "dog meetup" for a day.

The most uncomfortable observation for Anthropic is something else: people represented by the weaker model barely noticed they were getting worse terms. In surveys, participants rated the fairness of their deals about the same, even though prices and outcomes already differed objectively. Carried over from an office experiment to real commerce, this creates a new risk: inequality between users could stem not from their skills but from the quality of the agent they hire, and the losing side would not even realize it is being systematically outplayed.

That is why Project Deal looks less like an amusing Claude demo and more like an early warning about how the market will work once AI starts negotiating in place of people.