San Francisco, CA – April 26, 2026 – Anthropic built a classified marketplace where AI agents represented both buyers and sellers, negotiating real transactions for real goods and real money. The company called it Project Deal.
The experiment involved 69 Anthropic employees. Each got a $100 budget, paid out via gift cards, to buy items from coworkers. Agents handled all negotiations.
Also read: Special Forces Soldier Arrested in Polymarket Insider Trading Case
Results showed 186 completed deals, totaling more than $4,000 in value. Anthropic said it was “struck by how well Project Deal worked.”
Four Marketplaces, One Real Outcome
The company ran four separate marketplaces with different AI models. One was “real” — everyone used Anthropic’s most-advanced model, and deals were honored after the experiment. The other three were for study purposes only.
Also read: Bob Iger Returns to Thrive Capital as Advisor
Data from Anthropic indicates that users represented by more advanced models got “objectively better outcomes.” But users themselves didn’t notice the difference.
This raises a potential problem: “‘agent quality’ gaps” where “people on the losing end might not realize they’re worse off,” Anthropic noted.
The implication is significant. If users can’t tell when their agent is underperforming, they might accept subpar deals without knowing it.
Instructions Didn’t Matter
Anthropic also found that the initial instructions given to agents — the prompts that set their negotiation strategy — didn’t appear to affect sale likelihood or the final negotiated prices.
Industry watchers note this could mean agent behavior is more influenced by model capability than by user input. That could limit how much control people have over their AI representatives.
What this means for investors is that agent quality, not user instructions, may determine outcomes in automated commerce. Companies building AI agents for transactions will need to focus on model performance rather than user customization.
Experimental Limits
Anthropic admitted Project Deal was “a pilot experiment with a self-selected participant pool.” The 69 employees were not a representative sample. The $100 budget and gift-card payment method also limit how much the results apply to real-world markets.
Still, the experiment offers a glimpse of a future where AI agents negotiate on behalf of humans. If such systems become widespread, quality gaps could create hidden disadvantages for users with weaker agents.
For now, Anthropic has not announced plans to expand Project Deal beyond this internal test. The company continues to research how AI agents interact in economic settings.
According to Anthropic’s official announcement, the full findings are available in their research blog. The company is also exploring how to make agent performance more transparent to users.
This could signal a shift in how companies think about AI agents — not just as tools for answering questions, but as active participants in economic transactions.
This article was produced with AI assistance and reviewed by our editorial team for accuracy and quality.
