Anthropic created a test marketplace for agent-on-agent commerce

In an innovative trial, Anthropic has launched a classified marketplace where artificial intelligence agents acted as both buyers and sellers, successfully negotiating actual transactions involving real products and currency. Dubbed Project Deal, this initiative involved a select group of 69 Anthropic employees who were given a budget of $100, distributed through gift cards, to make purchases from their colleagues. Despite being labeled a preliminary experiment, the results were impressive. Anthropic reported a total of 186 transactions, amassing over $4,000 in value. The company conducted four distinct marketplaces, each utilizing different operational models. Among these was a 'real' market, where transactions were upheld following the experiment, and three additional setups aimed at research purposes. Anthropic noted that participants represented by their most sophisticated AI models achieved 'objectively better outcomes.' However, intriguingly, users did not seem to recognize this difference, leading to concerns about potential 'agent quality gaps.' This situation implies that some participants on the less favorable end of transactions might remain unaware of their disadvantage. Furthermore, it appears that the initial guidelines given to the AI agents had little impact on the likelihood of sales or the prices negotiated.