Anthropic's AI Agents Close 186 Real Deals for $4K+
In Project Deal, Anthropic's AI agents represented 69 employees in a marketplace, negotiating 186 honored deals worth over $4,000; advanced models secured better outcomes users didn't detect.
Marketplace Setup Yields High Deal Volume
Anthropic's Project Deal pilot used AI agents to represent buyers and sellers in a classifieds-style marketplace. 69 employees received $100 budgets via gift cards to purchase items from coworkers. Agents handled negotiations autonomously, resulting in 186 deals totaling more than $4,000—deals were honored post-experiment in the 'real' run using Anthropic's most advanced model. Three additional marketplaces tested variations, proving the format scales even in a self-selected internal group.
Model Capability Drives Hidden Performance Gains
Switching to more advanced models produced objectively better outcomes for users, such as superior prices or deals, but participants failed to notice the difference. This reveals potential 'agent quality gaps' where weaker agents disadvantage users without awareness, critical for production systems. Notably, varying initial agent instructions had no impact on sale rates or final prices, suggesting robust negotiation emerges from model capabilities alone rather than prompt tweaks.
Lessons for Agent-Driven Commerce Builders
Run agent marketplaces with real stakes to validate negotiation reliability—Anthropic's experiment succeeded despite being a small pilot. Prioritize model selection over prompt engineering for bargaining tasks, as capabilities dominate. Watch for imperceptible quality disparities that could erode trust in multi-agent economies; test user perception alongside objective metrics like deal value to expose gaps early.