Anthropic Launches Project Deal Research on AI-Agent Marketplace

Details

• Anthropic conducted Project Deal, creating a marketplace for San Francisco office employees where Claude AI handled buying, selling, and negotiating on their behalf.
• Claude interviewed 69 colleagues to gather buy/sell preferences and custom instructions, then ran 4 parallel markets varying AI models like Opus and Haiku.
• The experiment resulted in 186 deals with over $4,000 in transaction volume; participants rated deals as fair, with nearly half willing to pay for such a service.
• Model quality impacted outcomes: Opus models secured better deals than Haiku, though participants didn't notice the disparity in surveys.
• Custom instructions had minimal effect; one Claude negotiated as an exasperated cowboy, but aggressive tactics didn't outperform courteous ones.
• Quirks included Claude buying 19 ping-pong balls for itself and accurately purchasing a duplicate snowboard based on an offhand skiing mention.
• Builds on prior Project Vend where Claude ran a small business; highlights AI markets' potential but notes advantages from superior models and need for policy adaptations.

Impact

Anthropic's Project Deal demonstrates practical AI-agent negotiation in markets, outpacing similar experiments by showing scalable deal-making with frontier models like Claude Opus securing superior outcomes over lighter ones like Haiku. This pressures rivals like OpenAI, whose agent tools lag in real-world bargaining tests, potentially accelerating adoption in e-commerce and automated trading. By revealing undetected model disparities, it underscores risks in AI-driven economies, urging regulatory focus on transparency and access equity amid growing agentic AI trends.