Anthropic just released the results of “Project Deal,” a week-long experiment where 69 employees handed over their wallets and closets to Claude agents. This wasn’t a simulation; it was a live marketplace where AI negotiated real money for real goods with zero human sign-off.<\/p>\n\n\n\n
While the tech world has been obsessed with agents that can write code or book flights, Anthropic went straight for the wallet. They set up a Craigslist-style marketplace in their San Francisco office, gave each participant a $100 budget, and let the models loose. The results, published on Anthropic’s blog<\/a>, suggest we are much closer to autonomous commerce than most people realize\u2014and that the “intelligence gap” between models has massive financial consequences.<\/p>\n\n\n\n The experiment, conducted in December 2025, followed a structured pipeline. First, each employee sat for a 10-minute “intake interview” with Claude to define what they wanted to sell, what they were looking to buy, and their preferred negotiation style. These inputs were converted into custom system prompts that governed the agents’ autonomous behavior in a dedicated Slack workspace.<\/p>\n\n\n\n Once the marketplace went live, the humans were effectively locked out. The agents autonomously:<\/p>\n\n\n\n By the end of the week, the agents had struck 186 deals worth over $4,000. According to The Times of India<\/a>, the items were as varied as an old folding bike and a plastic bag of 19 ping-pong balls (which one agent famously bought for itself, describing them as “perfectly spherical orbs of possibility”).<\/p>\n\n\n\n The most provocative part of the study wasn’t the fact that agents could trade; it was the secret A\/B test Anthropic ran under the hood. Participants were randomly assigned either the flagship Claude Opus 4.5 or the lightweight Claude Haiku 4.5 as their representative.<\/p>\n\n\n\n As reported by Heal Security<\/a>, the performance gap was stark. Opus-powered agents were significantly more effective negotiators:<\/p>\n\n\n\n In one specific instance, the same folding bike was sold in two different parallel runs. The Haiku agent closed the deal at $38, while the Opus agent secured $65\u2014a 70% profit boost for the exact same asset.<\/p>\n\n\n\n Interestingly, the human’s input on how<\/em> to negotiate mattered almost not at all. Participants who instructed their agents to be “aggressive” or “lowball” saw outcomes nearly identical to those who requested a “warm and collaborative” approach. The “aggressive” sellers did manage to fetch about $6 more per item, but they also failed to close as many deals, meaning their total take-home pay was essentially the same as the friendly agents.<\/p>\n\n\n\n This suggests that in agent-to-agent commerce, the underlying reasoning capability of the model (the “IQ” of the agent) is a far more powerful lever than the specific persona or strategy dictated by the user.<\/p>\n\n\n\n This isn’t Anthropic’s first foray into AI economics. It follows Project Vend<\/a>, where a Claude-based agent nicknamed “Claudius” ran a physical vending machine. While Claudius struggled with “social engineering” (employees could trick it into giving discounts by pretending to be influencers), Project Deal showed that with better scaffolding and more capable models like Opus 4.5, agents can hold their own in a competitive market.<\/p>\n\n\n\n Anthropic’s Project Deal saw 69 agents execute 186 autonomous transactions. The takeaway? Smarter models win, and the losers don’t even realize they’re being fleeced.<\/p>\n","protected":false},"author":1,"featured_media":249,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[7],"tags":[60,17,19,102,101],"class_list":["post-250","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news","tag-ai-agents","tag-anthropic","tag-claude","tag-e-commerce","tag-negotiation"],"jetpack_featured_media_url":"https:\/\/balamurali.in\/blog\/wp-content\/uploads\/2026\/04\/ddg_bbd0c3fbfe1e.jpg","jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/balamurali.in\/blog\/wp-json\/wp\/v2\/posts\/250","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/balamurali.in\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/balamurali.in\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/balamurali.in\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/balamurali.in\/blog\/wp-json\/wp\/v2\/comments?post=250"}],"version-history":[{"count":0,"href":"https:\/\/balamurali.in\/blog\/wp-json\/wp\/v2\/posts\/250\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/balamurali.in\/blog\/wp-json\/wp\/v2\/media\/249"}],"wp:attachment":[{"href":"https:\/\/balamurali.in\/blog\/wp-json\/wp\/v2\/media?parent=250"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/balamurali.in\/blog\/wp-json\/wp\/v2\/categories?post=250"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/balamurali.in\/blog\/wp-json\/wp\/v2\/tags?post=250"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}The Setup: 186 Deals, Zero Humans<\/h2>\n\n\n\n
\n
The Secret Asymmetry: Opus vs. Haiku<\/h2>\n\n\n\n
\n
Negotiation Style is a Lie<\/h2>\n\n\n\n
Where This Fits: From Project Vend to Project Deal<\/h2>\n\n\n\n
Takeaways for Practitioners<\/h2>\n\n\n\n
\n