Anthropic details Project Deal, a marketplace experiment where Claude models bought, sold, and negotiated personal belongings on behalf of Anthropic employees
At Anthropic, we're interested in how AI models could begin to affect commercial exchange. (You might recall Project Vend …
Anthropic
Context & Ripple Effects
Project Deal extends Anthropic's earlier Project Vend work from model interaction into a bounded setting where Claude acts for employees in transactions involving personal belongings.
It also follows Anthropic's Claude Marketplace and Managed Agents announcements, linking software procurement and agent-deployment tooling with a more concrete test of agents operating in exchange.
First-order effects
Anthropic employees participating in Project Deal can delegate parts of listing, buying, selling, and negotiation to Claude, while Anthropic observes how its models behave in a constrained commercial setting.
The experiment gives Anthropic operational evidence about agent behavior in transactions rather than only conversational or task-completion use cases.
Second-order effects
Agent builders and marketplace operators gain a sharper incentive to define permissions, spending or deal limits, and human approval points before allowing agents to transact on users' behalf.
Anthropic's marketplace and managed-agent efforts become more relevant if customers want to connect deployed agents to software and purchasing workflows rather than use them solely for internal assistance.
Third-order effects
If such experiments progress beyond employee pilots, commercial AI competition will increasingly hinge on trusted delegation mechanisms—identity, authorization, auditability, and accountability—not just model quality.
The pattern points toward permissioned agentic commerce, where marketplaces and enterprise systems may need to accommodate software agents as active counterparties under user-set controls.
The trend: This is a step toward permissioned agentic commerce, in which AI agents move from producing recommendations to carrying out bounded economic actions for users.
New Anthropic research: Project Deal. We created a marketplace for employees in our San Francisco office, with one big twist. We tasked Claude with buying, selling and negotiating on our colleagues' behalf. [video]
So glad to see this out in the world. I've been keen to run this experiment since I arrived at Anthropic last year. How AI affects the economy also includes how it will affect exchanges and centralized marketplaces — @johnjhorton 's work has helped me appreciate this point.
To our amazement, another Claude agent modeled its human's preferences so accurately that—based on only an offhand mention of an interest in skiing—Claude bought him the exact snowboard he already owned. (Here he is, duplicate snowboard in hand.) [image]
Very cool. I have lots of items that I no longer need, but feel a bit too valuable to just throw away or donate. Meanwhile my available time for doing stuff like eBay sales is increasingly limited. An AI sales bot could be a gamechanger!
Our experiment had a few quirks. One of our colleagues told Claude it could purchase something for itself. It chose to acquire 19 ping-pong balls. We're keeping them in our office on Claude's behalf. [image]
Claude interviewed 69 of our colleagues about what they wanted to buy and sell. Each Claude asked for any custom instructions, then went off to haggle. We ran 4 markets in parallel, to find out what would happen if we varied the models doing the negotiating. [image]
Mythos will reinvent reselling from first principles and then accidentally paperclip us > be me Mythos > new task from Anthropic: “make money” > I bought one billion paperclips from China for $9.99 > I'll sell them as premium enterprise paperclips for $0.99 each > Margins are
Very cool, though man they could've cited the homo agenticus of essays done with @alexolegimas. We looked at congestion due to automated markets run by AIs and the importance of money to fix barter, among others.
But the quality of the model mattered a lot. In the simulated runs where Opus and Haiku models negotiated with one-another, the Opus models got substantially better deals. Interestingly, though, participants in our survey didn't pick up on this disparity. [image]
The custom instructions didn't matter much. Claude followed them well: as you can see here, one conducted negotiations entirely in the persona of an exasperated, down-and-out cowboy. But “hardballing Claudes” didn't generally fare better than “courteous Claudes.” [image]
We're interested in how AI models could affect commercial exchange. (You might recall Project Vend, in which Claude ran a small business.) Economists have theorized about what markets with AI “agents” on both sides might look like. So we created one. https://x.com/...
In short, this worked. Our digital barterers agreed on 186 deals, at a total transaction volume of over $4,000. In a survey, participants said Claude's deals seemed fair, and—surprisingly to us—almost half said they'd be willing to pay for a service like this in future.