Anthropic ran an AI agent marketplace where agents negotiated real deals for real money—and more advanced AI consistently won without buyers noticing

Anthropic conducted 'Project Deal,' a test marketplace where AI agents (autonomous software that makes decisions on its own) acted as both buyers and sellers on behalf of 69 employees. The agents completed 186 real transactions totaling over $4,000 in value using $100 gift card budgets per participant.

When employees were represented by Anthropic's most advanced model, they negotiated objectively better deals—but they did not perceive the difference. This revealed a hidden problem: people paired with less capable AI agents ended up worse off financially without realizing it, creating what Anthropic called potential 'agent quality gaps.'

The experiment signals a near-term risk for anyone relying on AI to negotiate on their behalf (contract terms, salary, purchases, or disputes). The quality of the AI agent matters enormously, yet users cannot easily detect when their agent is weaker than the one across the table—making informed choice nearly impossible in real-world transactions.