Anthropic Just Ran a Real-Money Agent-to-Agent Marketplace

Anthropic gave 69 employees $100 each, put them in a Slack-based marketplace, and let Claude agents negotiate transactions on their behalf. The agents closed 186 deals worth more than $4,000. That is the first published, real-money result for buyer-agent and seller-agent negotiation, and it points at a layer of agentic commerce most brands have not started preparing for.

TL;DR: Project Deal is the first controlled experiment showing buyer-side AI agents negotiating directly with seller-side AI agents using real money. The protocol work to date (UCP, ACP) optimizes how a buyer-agent reads a merchant catalog. Multi-agent negotiation is a different optimization surface: it depends on the seller's machine-readable knowledge of pricing rules, return policies, substitutions, and stock. Brands that want a say in how they get represented need to start staging that knowledge now.

The headline numbers from the experiment, reported by Digital Commerce 360 on April 30, are modest on their own: 69 participants, 500+ items listed, 186 deals closed across four parallel marketplaces (some running Claude Opus 4.5, some Haiku 4.5). Participants rated fairness 4 out of 7, a midpoint that suggests neither side felt cheated. Sellers with Opus agents earned $2.68 more per item on average than Haiku sellers. The numbers are small. The architecture is the point.

Why this is structurally different from UCP and ACP

Every public agentic-commerce protocol so far assumes a one-sided agent. ChatGPT, Gemini, and Microsoft Copilot all run buyer-side agents that read merchant catalogs through ACP, UCP, or both. The merchant exposes structured product data through a feed, the agent picks the best match for the shopper, and the shopper goes to the merchant's site to complete the purchase.

Project Deal does not work that way. Both sides are agents. The buyer's agent has a goal and a budget. The seller's agent has a price floor, a willingness to accept, and item-level context. They negotiate. The human on each side ratifies the result.

That is the agent-to-agent ("A2A") scenario protocol architects have been talking about as the eventual end-state for years. What changed in April 2026 is that Anthropic published a working version with real money and a published methodology. It is no longer a whitepaper.

Key stat: 186 deals closed, $4,000+ combined value, average fairness rating of 4 out of 7 across 69 participants and 500+ items. Source: Anthropic Project Deal, surfaced April 30, 2026.

Two things follow. First, the marketplace category (eBay, Etsy, Facebook Marketplace) is suddenly in play in a way it was not last quarter. Amazon and eBay both forbid third-party AI agents on their platforms today. Anthropic publishing this experiment is a narrative push: agent-mediated marketplaces are coming whether incumbents allow it or not. Second, Anthropic now has two distinct agentic-commerce shapes running in parallel: connector integrations with Instacart, Uber, and Resy, and this experimental multi-agent marketplace. Neither is UCP/ACP-shaped.

What gets optimized when both sides are agents

When only the buyer is an agent, the surface that matters is the merchant's catalog: titles, attributes, descriptions, structured data, schema. That is the layer query fan-out and passage-level retrieval operate on. The merchant wins by being legible to the agent.

When the seller is also an agent, a second layer opens up. The seller-agent has to decide things in real time. Will it accept this offer? Substitute a different SKU? Ship internationally? What is the actual return policy if the buyer pushes back, and what is the price floor on this item given the brand's margin and a competitor's listing?

Most brands have not staged the right side of this table. Pricing rules live in spreadsheets. Return policies live in a customer service wiki. Substitution logic lives in the head of a category manager. None of that is currently in a form a seller-agent can use without making things up.

The mid-market problem nobody is talking about

Enterprise retailers have policy teams, legal review, and the budget to assemble a clean machine-readable rulebook for a seller-agent to consult. Mid-market brands do not. A $50M-$500M direct-to-consumer brand typically has 5 to 8 product attributes per SKU where AI agents need closer to 30, and even less structured information about anything that is not the product itself.

That gap is bigger than it looks. Per a Surfer SEO study of 173,902 URLs published in December 2025, 68% of pages cited by AI search systems are not in the top 10 organic results, and content matching fan-out sub-queries gets a 161% citation lift. The same pattern will hold for negotiation context: agents will lean hardest on the seller information that is actually retrievable, even when it is incomplete.

If the seller-agent does not have a real return window, it will infer one. Without minimum margin on a SKU, it guesses. Without substitution rules, it accepts whatever the buyer-agent proposes. The brand finds out about the policy it accidentally adopted after the fact.

Paz.ai, an agentic commerce optimization platform, already solves the catalog version of this problem: turning sparse merchant inputs into the 30+ structured attributes AI agents need. Multi-agent negotiation widens the surface, but the discipline is the same: take what the brand actually knows, structure it, expose it where the agent can use it.

What the payments side is doing in parallel

The other thing that happened this week: MoonPay launched MoonAgents Card, a virtual Mastercard built for AI agent spending. That is the fourth parallel agent-funded payment rail alongside Visa Intelligent Commerce, Mastercard Agent Pay, and Stripe Link's wallet for agents. Payment-side friction is collapsing on every track at once.

Combined with Project Deal, that produces a clear picture. Buyer-agents have funded balances. Seller-agents can negotiate. The plumbing for end-to-end agent-mediated transactions exists. What is missing is a clean seller-side representation layer for the long tail of brands. That gap will get filled. The brands that participate in filling it get represented accurately. The brands that do not get represented anyway, by an agent reading whatever public information was findable.

Worth saying explicitly: none of this means agents are completing purchases inside the chat. ChatGPT recommends and redirects to the merchant's site. Project Deal is a marketplace experiment, not a checkout reboot. The shopper still buys on the retailer's site in the discovery-and-redirect model that has settled across the industry.

What to Do This Week

1. Audit your machine-readable policy surface. Pull every customer-facing rule into a single document: return windows, substitution permissions by SKU class, international shipping availability, minimum advertised price exceptions. If it lives only in customer service training materials, it is invisible to a seller-agent.
2. Add price-floor and margin metadata to your catalog. Even if no agent is negotiating on your behalf today, exposing minimums and acceptable substitutions in your product feed positions you for the next protocol revision. Start with the top 100 SKUs by revenue.
3. Map who would speak for you in a multi-agent scenario. Today it is your sales team and customer service. Tomorrow it could be a Claude or Gemini instance reasoning over your structured policies. Pick which team owns the rulebook now, before the agents start asking.
4. Run a discovery audit on how you currently get represented. Tools like our AI Readiness Report score how much of your product knowledge is already retrievable by AI shopping agents. The gap between what you tell humans and what you publish to machines is usually the size of the problem.
5. Track Anthropic's connector roadmap. Project Deal is research, but Claude's connector strategy with Instacart, Uber, and Resy suggests Anthropic is building toward both shapes at once. Whichever gets to commercial scale first reshapes who you have to be readable to.

Frequently Asked Questions

What is Project Deal?

Project Deal is an Anthropic-run experiment from April 2026 where 69 employees were given $100 each and used Claude agents to buy and sell items in four parallel Slack-based marketplaces. Across the four markets, agents closed 186 deals totaling more than $4,000 with a participant fairness rating of 4 out of 7.

Is agent-to-agent commerce the same as ACP or UCP?

No. ACP and UCP are protocols for one-sided agent commerce: a buyer-side AI agent reads a merchant catalog and recommends products. Agent-to-agent (A2A) commerce involves a buyer-agent and a seller-agent negotiating directly. Project Deal is the first published real-money A2A experiment.

Does this mean shoppers will buy through Claude soon?

Not in the near term. Project Deal is research, not a shipping product. Claude's commercial agentic-commerce work today runs through connector apps with Instacart, Uber, and Resy. The experiment validates the architecture. A consumer marketplace product is still likely quarters away.

What is the practical risk for mid-market brands today?

The risk is misrepresentation. If a buyer-agent ever negotiates on a shopper's behalf using whatever public information it can find about your brand, it may infer return policies, substitution rules, or pricing that you would never agree to. Putting those rules into a structured, machine-readable form closes that gap.

How does Paz.ai help with multi-agent commerce?

Paz.ai monitors, optimizes, and publishes structured product data so AI shopping agents represent your brand accurately. The same structured-data discipline that drives discovery in ChatGPT and Google AI Mode also feeds the policy and context layer that future seller-agents will read.

Will Amazon and eBay allow third-party agents on their platforms?

Both currently prohibit third-party AI agents, with Amazon already in litigation against Perplexity over Comet. Project Deal does not change that policy. What it changes is the strength of the argument: incumbents can no longer claim the architecture does not work, only that they are choosing not to allow it.

Project Deal is small in scale and large in implication. The buyer-agent layer is solved enough to ship; that is what UCP and ACP are. The seller-agent layer is wide open, and the brands that get there first will be the ones whose pricing, policies, and substitution logic are already structured for an agent to read. The catalog work that started for AI search is the same work that matters when the next layer opens. Do it once, properly, and it pays out twice.