Anthropic tested a market where AI agents negotiate and conclude deals themselves

Anthropic tested a market where AI agents trade with each other and conclude deals without human participation in each step, and it already doesn't look like science fiction. In the Project Deal experiment, the company created an internal marketplace where Claude represented both sellers and buyers, published lots itself, conducted negotiations, made counteroffers, and finalized deals. Importantly, this was not a simulation on test data: employees actually exchanged real things and money after the experiment concluded.

The experiment involved 69 Anthropic employees in the San Francisco office. Each received a conditional budget of $100, which was later compensated with gift cards based on their purchases and sales. Before launch, Claude conducted a short interview: found out what the person was willing to sell, what they wanted to buy, at what price, and how exactly the agent should conduct negotiations.

After that, a separate agent with an individual system prompt was created for each participant. Trading took place within Slack channels, and people didn't confirm each deal manually: once started, agents acted independently. Anthropic simultaneously launched four versions of the market.

One was considered "real" — based on its results, participants later exchanged goods; the other three were needed for comparison. In two runs, everyone was represented by Claude Opus 4.5, while in two others, participants were randomly divided between Opus 4.

5 and the lighter Haiku 4.5 model. In the "real" run, agents concluded 186 deals across more than 500 lots worth just over $4,000.

These were not instant one-click purchases: agents had to search for matching interests, dispute prices, respond to counteroffers, and bring the conversation to agreement. The main conclusion turned out not to be that AI coped in general, but that model quality directly influenced the result. According to Anthropic, users with Opus concluded on average about two more deals than users with Haiku.

When the same item was sold by Opus, the price on average turned out to be $3.64 higher, and in the overall assessment, Opus as a seller brought in about $2.68 extra, while as a buyer, conversely, reduced the price by approximately $2.

45. In individual examples, the difference looked even more noticeable: the same lab-grown ruby that an Opus agent sold for $65, Haiku sold for only $35; a broken folding bicycle went for $65 in one run versus $38 in another. Interestingly, the style of instructions almost didn't change the outcome.

Some employees asked their agents to be soft and friendly, others to negotiate hard and start with lowball offers. Statistically significant effect wasn't observed: aggressive prompts didn't increase the chance of a sale and didn't consistently help get a better price. However, other features of agent trading became apparent.

One agent bought a person an almost identical snowboard to one they already had, apparently taking the owner's tastes too literally. Another, at a female employee's request, chose a gift "for Claude himself" — a pack of 19 ping-pong balls for $3. There were also deals not about things but about experience: for example, agents agreed on a free "dog meetup" for the day.

The most unpleasant observation for Anthropic was different: people represented by the weaker model almost didn't notice they were getting worse terms. In surveys, participants rated the fairness of deals roughly equally, although objectively the difference in prices and outcomes was already there. If this moves from an office experiment to real commerce, a new risk emerges: inequality among users could arise not from their skills, but from the quality of the hired agent, and the losing side won't even realize they're being systematically outplayed.

Therefore, Project Deal looks not just like an amusing demonstration of Claude, but like an early warning about how the market will be organized when AI starts trading instead of people.