On Friday, Anthropic disclosed a strange internal experiment called Project Deal, a test marketplace where Claude-based AI agents negotiate with each other on behalf of real employees.
The company set up a Craigslist-style marketplace in its Slack workspace and let AI agents, not humans, handle all offers and counteroffers. Employees in the San Francisco office listed personal items and gave Claude guidance on what they wanted to buy or sell, their price limits, and their tolerance for negotiation.
In total, 69 employees participated in the experiment, each receiving a budget of around $100 (often via gift cards). The agents handled more than 500 unique items and completed 186 transactions totaling just over $4,000. After the experiment, participants carried out the deals in real life by exchanging money and goods according to what their agents agreed to during the marketplace runs.
Anthropic ran four parallel versions of the marketplace. One used Claude Opus 4.5 agents, while the others used lightweight models, like Claude Haiku, to represent buyers and sellers. Stronger models reportedly performed better at negotiation, enabling users with more capable agents to secure higher-quality deals.
Anthropic positions Project Deal as an early probe of probable future AI-to-AI marketplaces, not a public product. Though a bit quirky, the experiment isn’t the first bizarre project Claude has given its LLMs; in March, its retired model began writing its own blog.
