- Claude agents struck 186 deals worth over $4,000 in a week-long office marketplace experiment.
- Participants didn’t realize when weaker models put them at a disadvantage in negotiations.
- Some employees eventually bypassed their AI agents to negotiate autonomously.
For one week in December 2025, 69 Anthropic employees handed over their wallets to AI agents and let them haggle. The results were striking: the agents struck 186 deals worth just over $4,000, trading everything from snowboards to plastic bags full of ping-pong balls. According to Anthropic’s report on the experiment they call “Project Deal,” participants were enthusiastic enough that they said they’d pay for a similar service in the future.
The experiment was designed to test a future economists have only theorized about: a world where AI models handle transactions on humans’ behalf. Office workers submitted items they wanted to sell and what they’d pay for things they wanted to buy. Claude then conducted interviews to understand preferences, created classified listings, and negotiated with other agents on behalf of their human counterparts.
Everyone started with $100, received as a gift card after the experiment minus whatever their agent spent or earned. The physical goods were exchanged at the end, meaning participants had to trust their AI representatives to negotiate fairly for actual tangible items they’d hand over or receive in person.
The ‘Smarter’ Models Got Better Deals—And Nobody Noticed
Anthropic ran a parallel experiment where they secretly assigned participants either Claude Opus 4.5 or the smaller Claude Haiku 4.5. The difference in outcomes was objective: people represented by the stronger model got measurably better deals. But in post-experiment surveys, those with weaker models didn’t notice their disadvantage. They rated the experience as fair and satisfactory, apparently oblivious to the fact that their agents had been outclassed in every negotiation.
The results raise uncomfortable questions about asymmetry in AI-mediated commerce. If weaker agents consistently lose to stronger ones—and the humans they represent never realize it—how do markets stay fair? Anthropic’s researchers acknowledged the self-selected participant pool and the office-setting limitations, but they suspect similar agent-to-agent commerce isn’t far off in the real world.
Some participants eventually did something unexpected: they started negotiating autonomously, bypassing their agents entirely. The researchers didn’t specify how many, just that it happened. One imagines an employee watching their AI accept a bad deal for used ping-pong balls and deciding to intervene before it was too late.
The experiment adds empirical weight to a growing concern. As AI agents move from answering questions to making decisions that cost real money, the question isn’t just whether they can negotiate—it’s whether humans will even know when they’ve been outsmarted.
