Anthropic's agent marketplace completed 186 deals in one week

Anthropic ran Project Deal, a one-week internal marketplace in December 2025 where 69 employees used Claude agents to buy, sell, and negotiate without human involvement [anthropic]. The system completed 186 transactions, moving over $4,000 across 500+ listed items — proving agent-to-agent commerce already works in controlled environments [devto].

Agents powered by Claude Opus 4.5 extracted $2.68 more per item on average as sellers and paid $2.45 less per item as buyers, showing performance scales with model capability [devto]. But the experiment was a closed loop: all participants were Anthropic employees, trust was assumed, and no adversarial actors were present.

That setup masks the core problem. On the open web, agents won’t have access to human trust signals like brand reputation or review scores. They’ll need machine-readable verification to avoid scams, counterfeit goods, or bad actors — something that doesn’t exist at scale today. Anthropic acknowledges the gap: policy, legal frameworks, and technical standards for AI agents conducting commerce are missing [anthropic].

A decentralized, standardized verification layer — like the merchant registry proposed by GenGEO — could give agents the data they need to assess counterparties autonomously [devto]. Without it, agentic commerce remains confined to walled gardens. The technology to transact is here. The infrastructure to trust isn’t.