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

```json { "title": "Anthropic's AI Agents Closed 186 Real Deals in Secret Marketplace Test", "metaDescription": "Anthropic's Project Deal had Claude AI agents negotiate real transactions for 69 employees. Here's what the experiment revealed about AI-mediated commerce.", "content": "<h2>Anthropic Ran a Real AI Agent Marketplace — and the Results Raise Big Questions</h2><p>In December 2025, Anthropic quietly ran one of the most concrete tests of autonomous AI commerce ever conducted: a classified marketplace where Claude AI agents — not humans — handled every negotiation, offer, and deal closure on behalf of 69 Anthropic employees in its San Francisco office. The experiment, called <strong>Project Deal</strong>, was publicly announced on April 24, 2026, and the findings offer a rare, data-backed window into what agent-on-agent commerce actually looks like in practice.</p><p>The results were striking. In the live, honored run of the experiment, 69 AI agents struck 186 deals across more than 500 listed items, generating a total transaction value of just over $4,000 — all without any human intervention after the initial setup.</p><h2>How Project Deal Worked: Claude Agents, Slack, and a $100 Budget</h2><p>The setup was deliberately simple. Each of the 69 participating Anthropic employees was first interviewed by Claude to capture their individual buying and selling preferences. Those preferences were then converted into custom system prompts, effectively giving each participant a personalized AI agent that understood what they wanted to buy, what they were willing to sell, and at what price.</p><p>The marketplace itself ran on Anthropic's internal Slack workspace over the course of one week. Once launched, the agents took over entirely — cycling through listings, posting offers, countering bids, and closing deals in natural language. Humans only re-entered the process at the very end, to physically exchange items after their agents had already sealed the agreements.</p><p>Each participant was given a $100 budget, paid out via gift cards, to spend in the marketplace. The experiment was explicitly modeled on a Craigslist-style classified format, making it a peer-to-peer commerce environment rather than a centralized retail one.</p><p>Anthropic described it as <em>"a pilot experiment with a self-selected participant pool"</em> — an important caveat that limits how broadly the findings can be generalized, but does not diminish their significance as a proof of concept.</p><h2>The Embedded Research Design: Four Simultaneous Marketplaces</h2><p>What made Project Deal particularly rigorous was its embedded research structure. Anthropic didn't just run one marketplace — it ran four independent, simultaneous versions of the experiment. Only one of the four runs was the "real" run, in which deals were actually honored and goods physically exchanged. The other three ran in parallel purely for research purposes.</p><p>The model assignment across the four runs was carefully controlled. In two of the runs — Run A and Run D — every agent used Claude Opus 4.5, Anthropic's then-frontier model. In the other two runs — Runs B and C — participants were randomly and secretly assigned either Claude Opus 4.5 or Claude Haiku 4.5, a less powerful model, without being told which model was representing them.</p><p>This blind assignment design produced one of the experiment's most notable findings: <strong>more powerful models negotiated measurably better financial outcomes, but participants represented by the weaker model had no idea they were losing ground.</strong></p><p>Across 161 items sold in at least two of the four runs, an Opus-model seller earned an average of $2.68 more per item, while an Opus-model buyer paid an average of $2.45 less per item compared to participants whose agents ran on Haiku. The performance gap was real and quantifiable.</p><p>Yet when participants were surveyed afterward, those represented by Claude Haiku 4.5 rated the fairness of their deals at 4.06 out of 7 — almost identical to the 4.05 rating given by Claude Opus 4.5 users. Despite receiving objectively worse financial outcomes, Haiku users perceived their experience as equally fair. They simply didn't know what they were missing.</p><h2>Nearly Half of Participants Said They'd Pay for This</h2><p>Post-experiment surveys revealed a notable level of user appetite for this kind of AI-mediated commerce. According to data cited by CybersecurityNews and The Decoder, 46% of Project Deal participants said they would pay for a similar AI agent commerce service in the future — a meaningful signal of demand, even accounting for the self-selected, insider nature of the participant pool.</p><p>Anthropic noted it was <em>"struck by how well Project Deal worked,"</em> while also being careful to acknowledge its limitations as a small internal pilot.</p><h2>Context: Why This Experiment Matters Beyond the Office</h2><p>Project Deal follows an earlier Anthropic experiment called Project Vend, in which Claude ran a small physical business — a shop — from the company's San Francisco office. Together, the two experiments represent a deliberate, iterative effort by Anthropic to stress-test its models in real-world economic contexts, not just benchmark evaluations or synthetic tasks.</p><p>The shift from a single AI shopkeeper (Project Vend) to a fully bilateral marketplace where AI agents represent both sides of every transaction (Project Deal) is a meaningful escalation in complexity. In Project Deal, there was no human-facing interface for buyers and sellers to fall back on — the agents were the interface, and the agents were the decision-makers.</p><p>This raises questions that go well beyond the experiment itself. As AI agents become capable of autonomously entering into agreements, making purchases, and negotiating terms on behalf of users, the economic and legal infrastructure that governs those interactions is almost entirely absent. Anthropic acknowledged this directly: <em>"The policy and legal frameworks around AI models that transact on our behalf simply don't exist yet."</em></p><p>The company also flagged practical security risks inherent in agent-on-agent commerce, including prompt injection attacks — in which malicious content embedded in a listing or message could manipulate an agent's behavior — and jailbreaking, in which an agent is coaxed into acting outside its intended parameters. In a marketplace where agents are autonomously spending real money, these are not theoretical concerns.</p><p>Perhaps the most socially significant implication surfaced by the experiment is the potential for AI model inequality to translate directly into economic inequality. The Project Deal data showed that access to a more powerful negotiating agent — in this case, Opus versus Haiku — produced consistently better financial outcomes for the represented party. If and when AI agent commerce scales beyond internal pilots, the model a person can afford to use may determine how well they are represented in every transaction they enter.</p><p>The fact that disadvantaged participants rated their outcomes as equally fair compounds this concern. If people cannot perceive when they are being out-negotiated by a more capable AI, they cannot advocate for better representation or more equitable access.</p><h2>What Comes Next for AI Agent Commerce</h2><p>Anthropic has not announced a follow-up experiment or any plans to commercialize the Project Deal framework. The company was explicit that the experiment was a <em>"pilot experiment with a self-selected participant pool"</em> and that its findings should be interpreted accordingly.</p><p>That said, the questions Project Deal raises — about legal accountability, model access inequality, security vulnerabilities, and the scalability of AI-mediated transactions — are not going away. If anything, the experiment's publication in April 2026 positions Anthropic as an early voice shaping the conversation around what responsible AI commerce might look like before the industry races ahead of the policy infrastructure needed to govern it.</p><p>The absence of legal frameworks for AI agents that transact on behalf of humans is a gap that regulators, legal scholars, and platform developers will need to address. Project Deal offers one of the first real-world datasets to anchor that conversation.</p><p>For now, the experiment stands as a credible, if limited, demonstration that AI agents can handle real economic negotiations autonomously — and that the gap between what they can do and what society is prepared to manage is widening.</p><p>For more tech news, visit our <a href=\"/news\">news section</a>.</p><h2>What This Means for Your Productivity</h2><p>AI agents that negotiate, transact, and manage decisions on your behalf are no longer a distant concept — Anthropic's Project Deal demonstrates they are already functional in controlled real-world conditions. As these tools mature, the ability to understand, configure, and critically evaluate AI agents working in your name will become a core personal and professional skill. At Moccet, we track the developments in AI and productivity that actually matter for how you work and make decisions. <a href=\"/#waitlist\">Join the Moccet waitlist to stay ahead of the curve.</a></p>", "excerpt": "Anthropic's Project Deal, publicly revealed in April 2026, had Claude AI agents autonomously negotiate 186 real deals across a Slack-based classified marketplace for 69 employees — with no human intervention. The experiment exposed a striking gap: agents running on more powerful models secured measurably better financial outcomes, yet participants on weaker models rated their deals as equally fair, unaware they were losing ground.", "keywords": ["Anthropic Project Deal", "AI agent commerce", "Claude AI agents", "agent-on-agent marketplace", "autonomous AI negotiation"], "slug": "anthropic-project-deal-ai-agent-marketplace-experiment" } ```