Anthropic Tests Marketplace Where AI Agents Negotiate Deals
Anthropic’s ‘Project Deal’ tested whether AI agents could buy, sell and negotiate for employees, completing 186 transactions worth more than $4,000.
Topics
News
- AI Skills May Shape Appraisals Over Next Three Years
- Anthropic Tests Marketplace Where AI Agents Negotiate Deals
- AWS Reaffirms $12.7 Billion India Cloud Push
- DeepSeek Previews V4 Model for Agentic AI Tasks
- Reliance Meta AI Venture Names Parminder Singh CEO
- Meta Layoffs, Microsoft Buyouts Signal AI Reset
Anthropic has run an internal experiment in which artificial intelligence agents acted as buyers and sellers in a classified marketplace, negotiating and closing transactions for human users.
The week-long pilot, called Project Deal, involved 69 employees at Anthropic’s San Francisco office. Each participant was represented by an AI agent and given a $100 budget to buy and sell personal belongings.
“For one week, we created a classified marketplace for employees in our San Francisco office, like Craigslist, but with a twist all of the deals were conducted by AI models acting on our employees’ behalf,” Anthropic said in its official post.
The agents completed 186 transactions with a total value of more than $4,000. The goods ranged from household items to sports equipment, with physical handovers completed after the AI-negotiated agreements were finalized.
Anthropic said it had expected the test to expose limits in how AI agents understood user preferences and negotiated terms, but found the system worked better than anticipated.
“We were struck by how well Project Deal worked,” the company said.
Alongside the visible pilot, Anthropic ran a parallel test in which some participants were represented by more advanced models and others by smaller models. The company compared Claude Opus 4.5 with Claude Haiku 4.5.
Users represented by stronger models secured better outcomes, Anthropic said.
“We found that agent quality does make a difference: people represented by ‘smarter’ models got objectively better outcomes,” the company said.
But users represented by weaker models generally did not realize they had done worse.
“Our post-experiment survey found that those with weaker models didn’t notice their disadvantage,” Anthropic said.
The company said the experiment was limited by its small, self-selected internal participant pool, but argued that it offered an early view of how agent-to-agent commerce could develop.
“To be sure, this was a pilot experiment with a self-selected participant pool. But we suspect we’re not far from more agent-to-agent commerce bubbling up in the real world, with real consequences,” Anthropic said.
