Anthropic's Project Deal: AI Agents Negotiate Real Trades Between Employees
Summary
- • Anthropic ran Project Deal, a pilot marketplace where AI agents negotiated real purchases between 69 employees
- • 186 deals completed totaling over $4,000 across four separate test marketplaces with different model tiers
- • Users represented by more advanced models got objectively better outcomes — without the disadvantaged party noticing
- • Initial agent instructions had no measurable effect on sale likelihood or negotiated prices
Details
Anthropic ran Project Deal, a real-money agent-to-agent marketplace pilot with 69 employees.
Employees were each given a $100 gift card budget to buy and sell goods with coworkers, with AI agents acting as their representatives. Four separate marketplace variants were run simultaneously — one 'real' market where deals were honored post-experiment, and three for research purposes.
186 deals completed, totaling more than $4,000 in value.
This occurred within a self-selected pool of 69 Anthropic employees, making it a constrained but functional proof of concept. Anthropic described itself as 'struck by how well Project Deal worked,' indicating results exceeded internal expectations.
Users paired with more advanced models achieved objectively better deal outcomes.
Anthropic tested different model tiers across its four marketplaces and found a measurable performance gap favoring users whose agents ran on more capable models. The company did not specify which models were used in each variant.
Disadvantaged users did not perceive they were getting worse outcomes, raising an 'agent quality gap' concern.
Anthropic flagged this as a serious potential issue: in real-world deployments, users with access to less capable agents could consistently lose value in negotiations without any awareness of the disparity. This has implications for fairness, consumer protection, and market transparency in agentic commerce.
Initial instructions given to agents had no significant effect on sale likelihood or final prices.
Regardless of how users prompted or configured their agents at the outset, negotiation outcomes were not meaningfully influenced. This suggests that in competitive agent-to-agent settings, base model capability dominates over user-level customization.
Product Launch = experiment or tool release, Stat = quantitative result, Research = experimental finding, Insight = interpreted implication or risk
What This Means
Project Deal is an early but real demonstration that AI agents can act as economic proxies — negotiating, agreeing, and completing transactions on behalf of humans without step-by-step oversight. The fairness risk Anthropic identified is significant: if agent capability determines deal quality and users cannot perceive the gap, agentic commerce could quietly entrench inequality between those with access to premium models and those without. As agent-to-agent transactions scale beyond controlled experiments, the question of model-tier disclosure and outcome transparency will become a policy and market design problem, not just a research one.