Anthropic recently ran an experiment where the San Francisco-based AI startup noted how its AI models negotiated and completed real transactions on behalf of users in a controlled internal marketplace. Anthropic noted in a blog post that the experiment called “Project Deal” was a week-long pilot that was conducted at its San Francisco office, where employees used AI agents based on Claude to buy and sell goods without human intervention during negotiations.
How did Anthropic’s AI agent marketplace work?
Anthropic says it recruited 69 employees in December last year for the project and gave each of them a $100 budget to spend in a custom Slack-based marketplace. First, Claude interviewed the participants to determine what personal belongings they wanted to sell, what they wanted to buy, and their preferred negotiation style.
Based on these responses, Anthropic then assigned each participant a customised Claude AI agent with specific instructions on pricing and behaviour. These agents were then deployed into multiple Slack channels, where they independently posted listings, discovered potential matches, made offers, negotiated prices, and finalised deals. Anthropic says there was no human intervention once the marketplace went live, with agents not seeking approval from users during negotiations or during bidding.
Once agents reached agreements, they generated deal terms and confirmed trades. These were later executed in the real world, with employees meeting in person to exchange the goods that had been negotiated entirely by their AI representatives.
Anthropic says it was highly impressed with how the experiment went, with 186 deals being struck across more than 500 listed items, generating a total transaction value of just over $4,000.
“To our surprise, participants were very enthusiastic about the experience—they even stated a willingness to pay for a similar service in the future.”
Anthropic reveals the risk of AI inequality:
Anthropic says it ran four parallel versions of the marketplace to test how the participants fared if they were represented by different classes of Claude models. Out of the four experiments, there was one ‘real’ experiment where actual items changed hands and three other experiments that were only for its study. However, participants were not told which one until after the experiment concluded.
Two of these experiments used the then-flagship Opus 4.5 model, while the other two experiments used a mix of Opus and the smaller Claude Haiku 4.5 model.
Anthropic said the results revealed clear differences in outcomes depending on the model representing the user. Participants assigned to Opus agents completed more deals on average and were able to secure better prices in negotiations. In cases where the same item was sold across different runs, the Opus agent often managed to sell it for a higher price compared to Haiku.
Here’s what Anthropic’s data revealed:
The company then went on to warn about the ‘agent quality’ implications in real-world marketplaces, where differences in AI capability could create unequal advantages without users realising they are at a disadvantage.
“There was clearly a quantitative disadvantage to being represented by Haiku: these users got worse deals. But they didn’t seem to notice it. This has an uncomfortable implication: if “agent quality” gaps were to arise in real-world markets—and there is no reason to think they won’t—then people on the losing end might not realize they’re worse off,” the company said in its blog post.
Source link
#Anthropic #secretly #created #Claudepowered #marketplace #agents #bought #sold #goods #humans #Mint
