Analysis
Anthropic has published a GitHub blueprint for building shopping agents on Claude, giving developers a reference implementation that combines the Messages API, the Agent SDK and Claude managed agents with product catalog lookup, cart state, checkout and a store of customer preferences and purchase history, The Register reported. The canonical example: "I need a tent, sleeping bag, and stove for a weekend trip with two kids."
This is infrastructure, not a consumer product. Anthropic is not launching a shopping app; it is lowering the cost for retailers and marketplaces to build one on Claude, which is the same play it ran with the Model Context Protocol -- publish the plumbing, let the ecosystem standardize on it, sell tokens.
A Crowded Field
The competitive context is dense. OpenAI shipped Instant Checkout with Etsy and Shopify merchants and co-authored the Agentic Commerce Protocol with Stripe. Google has pushed AP2, its agent payments protocol, with card networks and more than 60 partners. Perplexity has run Buy with Pro since 2024. Visa and Mastercard have both published agentic payment credentials designed to authenticate a bot acting for a specific cardholder. Anthropic is the last of the major labs to put a stake in commerce, and it is entering through developers rather than consumers.
The demand side is the problem. Gartner found this year that only 11 percent of consumers are willing to let AI make purchase decisions for them. Accenture put the share willing to delegate purchases at 32 percent, with 74 percent comfortable letting agents handle routine tasks but just 9 percent open to fully autonomous shopping. Those numbers describe a market that wants an assistant, not a buyer -- research, comparison and cart-building, with a human pressing the final button.
Liability and Pricing Are Unresolved
The unresolved mechanics are liability and pricing. Monica Eaton of Chargebacks911 has pointed out that merchants have no framework for disputes when an agent buys the wrong thing: card network chargeback rules assume a cardholder, not a delegate. The Brookings Institution has warned that agentic shopping accelerates personalized pricing, because an agent negotiating on a consumer's behalf discloses far more about willingness to pay than a human clicking through a site does. Neither issue has a standard answer, and both fall on merchants first.
For retailers, the strategic question is whether to expose catalog and checkout to agents at all. Doing so risks commoditizing the storefront -- if agents compare across merchants on price and availability, brand and merchandising lose their leverage, which is precisely what happened to hotels under OTA distribution. Refusing risks invisibility if consumer behavior shifts. Most large retailers are hedging with limited, authenticated agent access rather than open APIs.
For founders, the opening is in the boring middle: agent-aware fraud scoring, dispute handling for delegated purchases, merchant-side policy engines that decide which agents can transact and under what limits. Pulse has tracked Anthropic's enterprise push through its API and agent releases, and commerce is the first area where its tooling arrives after the market has already picked competing protocols.
The adoption metric that matters is not agent traffic. It is completed transactions where no human reviewed the cart, and no lab has published that number.
