Anthropic Built a Test Marketplace Where AI Agents Trade With Each Other
Anthropic ran a closed pilot called Project Deal where AI agents bought and sold from each other on behalf of 69 employees, settling 186 trades and over $4,000 in volume across four model setups.
Anthropic has revealed an internal experiment called Project Deal, a classified marketplace where AI agents handled real purchases on behalf of human employees, offering one of the first structured glimpses at how agent-to-agent commerce might actually behave.
Inside Project Deal
According to TechCrunch, 69 self-selected Anthropic employees were each given a $100 gift-card budget and represented by AI agents that negotiated to buy items from coworkers also represented by agents. The company stood up four parallel marketplaces using different underlying models to compare outcomes, and the pilot closed 186 deals totaling more than $4,000 in transacted value.
The "agent quality gap"
Anthropic's takeaway was unusually candid: "when users are represented by more advanced models, they get objectively better outcomes," yet participants "did not always recognize the disparity." The company warned this could create "agent quality gaps" in the wider economy, where wealthier or more sophisticated buyers quietly benefit from better representation in the same way they do with lawyers or brokers today.
BlockAI News' View
Project Deal is small, but the framing is the story: Anthropic is treating agent commerce as something that needs benchmarking and guardrails, not just demos. For Web3, that legitimizes the thesis that programmable money rails — the kind already shipping on Base, Solana and L2 stablecoin networks — will be where this actually settles. Expect crypto teams to argue, loudly, that "agent quality gaps" are easier to audit when the venue is on-chain.
Want every AI × Web3 signal the moment it breaks? Subscribe to the BlockAI News daily brief.
