Claude agents haggled for Anthropic employees — and bought 19 ping-pong balls

Anthropic just ran an experiment that sounds like a fever dream from the future of commerce. The company created a marketplace for its San Francisco employees, then let Claude agents do all the buying, selling, and negotiating. They called it Project Deal.

Why do this? Anthropic wanted to see what happens when both sides of a transaction are AI agents. The company said it was curious about "how AI models might affect commercial transactions." It's a follow-up to Project Vend, where Claude ran a vending machine business.

Here's how it worked: Humans were involved only at the start and the end. Each participant told their assigned Claude agent what to sell or buy, their target price, and their negotiation style. Then the AI took over. Agents were randomly assigned either Claude Opus (Anthropic's high-end model) or Claude Haiku (the lightweight, cheaper one).

Anthropic set up four different markets, varying by which models participated and whether the humans could watch:

• Opus only, visible to participants.
• Opus only, not visible.
• Opus and Haiku, visible.
• Opus and Haiku, not visible.

The agents negotiated among themselves in a Slack channel. Once a deal was struck, the human employees met up in real life to exchange the goods.

By the end, employees reached agreement on 186 transactions totaling over $4,000. Participants said they felt the AI-run deals were fair, and nearly half said they'd pay for a service like this in the future.

The big finding? Model quality mattered a lot. When Opus and Haiku went head-to-head, Opus consistently got better terms — and humans had no idea. As Anthropic put it, "AI model quality was very important in transactions. In simulations where Opus and Haiku negotiated with each other, Opus got substantially better deals. Interestingly, participants in our study did not notice this difference." For example, Opus sold a similar bicycle for $38 more than Haiku did.

Custom instructions didn't make much difference to the outcome, though Claude followed them diligently. Tell it to act like "a grumpy, washed-up cowboy" and it would. But polite personas generally closed better deals than stubborn ones.

And then there's the WTF moment: One participant told their Claude it could buy something for them. The agent came back with 19 ping-pong balls.

Anthropic acknowledges the whole thing is rough around the edges. "Access to higher quality models gives a real advantage, but participants were unaware of it," the company said. "There are many other ways this could fail. Policy and legal frameworks need to adapt to keep up with this technology."