The Dawn of Agent-on-Agent Commerce
In a significant stride towards autonomous commerce, Anthropic recently unveiled the results of "Project Deal," an internal experiment that saw AI agents representing both buyers and sellers in a classified marketplace. This groundbreaking test, conducted in December 2025, involved 69 Anthropic employees in their San Francisco office, where AI agents powered by Claude models negotiated and finalized 186 real transactions for physical items, accumulating a total transaction value exceeding $4,000. These were not simple, one-click purchases; agents engaged in multi-turn negotiations, demonstrating contextual reasoning and personalization.
The experiment was set up within Anthropic's Slack workspace, mimicking a Craigslist-style marketplace. Each participating employee was first interviewed by a Claude AI agent to ascertain their selling preferences, buying wish lists, and desired negotiation styles. These inputs were then transformed into custom system prompts, allowing the agents to operate independently, posting listings, making offers, counteroffers, and ultimately closing deals without any human intervention during the negotiation process. Humans only re-entered the picture at the end to physically exchange the goods their AI agents had agreed to trade.
Unveiling Model Asymmetries and Performance Gaps
Beneath the surface of this successful demonstration, Anthropic conducted a parallel, secret experiment to assess the impact of different AI model capabilities. Participants were unknowingly assigned either the then-flagship Claude Opus 4.5 or the smaller Claude Haiku 4.5 as their negotiating agent. The results exposed a measurable and significant performance gap: Opus agents consistently achieved better outcomes, with sellers earning an average of $2.68 more per item and buyers saving $2.45 per item. Opus users also completed approximately 2.07 more deals overall.
Despite these demonstrable differences in outcomes, participants represented by the less powerful Haiku agents did not perceive any disadvantage, rating the fairness of their deals similarly to those with superior agents. This "invisible arms race" highlights a critical challenge for the future of AI agent commerce: the potential for performance disparities between models to create uneven negotiating power and subtle exploitation in peer-to-peer trades. Anthropic warns that this "uncomfortable implication" will require careful consideration by industry, regulators, and users before agentic commerce becomes mainstream.
The Path Forward for AI-Mediated Marketplaces
The success of "Project Deal" suggests that AI agents can indeed effectively represent humans in a marketplace, with 46% of participants expressing a willingness to pay for similar AI negotiation services in the future. This enthusiasm points towards the imminent emergence of commercial platforms integrating Claude-like agents, potentially within existing marketplaces such as eBay or Craigslist by late 2026. However, the experiment also underscores the need for developers to address model disparities, perhaps by equalizing capabilities or mandating disclosures to prevent imbalances from eroding trust in automated commerce.
The broader landscape of AI agent marketplaces is rapidly evolving. Companies like Google Cloud and Microsoft have already launched their own AI agent marketplaces, allowing developers and businesses to list, buy, and sell AI agents. Anthropic itself has also introduced a Claude Marketplace for B2B applications, positioning its platform as a distribution channel for enterprise software built on its AI models. As AI-driven trades involving real money and goods become more prevalent, regulatory scrutiny is expected to increase, focusing on fraud prevention and ensuring fairness in agent interactions.