On September 2, Anthropic published Building Commerce Agents with Claude on the Claude Blog. It is not another concept note about “agents that can shop.” It ships a runnable blueprint for a shopping agent and a merchant agent, vertical demos, a Claude Code plugin, and a partner narrative with Visa, Mastercard, Accenture, Shopify, and others—timed for holiday-season planning.
The post opens with customer-side results Anthropic attributes to retailers running shopping agents on Claude: carts up to about 35% larger and shoppers about 60% more likely to complete a purchase. Those figures are official customer claims without an attached third-party audit; the analysis below treats them as vendor-reported, not independently reproduced.
What the official post says
The core move is an engineering handoff. The repository is named as github.com/anthropics/commerce-agents, with full reference implementations of a shopping agent and a merchant agent that can run via the Messages API, the Agent SDK, or Claude Managed Agents (beta). Deploy targets include the Claude API, Amazon Bedrock, Microsoft Foundry, and Google Cloud Vertex AI. Solutions partners such as Accenture, Mastercard, and Visa are described as helping clients and merchant communities adopt the blueprints. Live vertical demos and an engineering deep-dive point to the August 28 post A guide to the anatomy of effective commerce agents.
The shopping agent lives inside an app or website. The blueprint supplies integration points for catalog, cart, checkout, preferences, and order history, while leaving payment to the merchant’s existing checkout or an agentic payments provider. In the official example—a customer asking for a tent, sleeping bag, and stove for a weekend trip with two kids—the agent should search the catalog, assemble multi-item sets, remember preferences, show products and the cart in-conversation, and answer shipping, return, and refund questions without bouncing the shopper to a help center. Guardrails are described as constraining prices and products to real catalog data and avoiding manipulative upsell patterns, implemented in-repo as skills and tools for catalog search, multi-item planning, deep research, personalization, customer care, and in-conversation UI.
The merchant agent supports store operators: answering what to discount to clear last season’s inventory from the store’s own data, tracking inventory and flagging SKUs that may sell out before a promotion, recommending pricing and promotions from sales history, and drafting campaigns. Anthropic stresses that when the agent proposes a change, a person must approve it before anything goes live. Matching skills cover sales analytics, catalog and inventory, marketing and promotions, and in-portal charts and dashboards.
Customer quotes span Visa, Mastercard, Accenture, Priceline, Intuit, Shopify, Klaviyo, Wix, Zomato, Square, and others. Checkable engineering anecdotes include Wix saying engineers had a commerce agent taking prompts within about fifteen minutes, and another team saying both agents ran locally in well under an hour, with Claude Code producing two different architectures across two runs. Accenture cites its research that 85% of people are open to collaborating with an AI agent and nearly three in four would trust a personal AI agent more than their best friend to buy on their behalf—partner research, not an Anthropic bench.
Product and technical value
Turning “a chatty shopping guide” into a forkable repo compresses trial-and-error. Encoding catalog constraints, in-conversation UI, tool-iteration limits, and prompt caching into the default scaffold admits that commercialization bottlenecks are no longer whether the model can talk, but whether checkout, inventory truth, brand voice, and compliance boundaries can be encoded in one runtime. Multi-item planning plus same-thread support on the shopper side, and suggest-then-human-approve on the merchant side, package a procurable division of labor: agents for throughput and watchfulness, humans for pricing power and brand risk.
For developers, three entry points—Messages API, Agent SDK, Managed Agents—reduce the cost of piloting without a stack rewrite. For retail and travel platforms, shipping vertical demos before the holidays compresses the sales cycle into the buying window. Pointing September 2’s delivery post at August’s anatomy essay keeps roles clean: one is distribution and partner network; the other is design doctrine. They should not be flattened into the same rewritten story.
Competition and strategy
Agentic commerce is becoming a contested intersection of payment networks, commerce SaaS, and foundation-model vendors. OpenAI and peers are also pushing buy-in-conversation flows; Google and Amazon hold distribution and catalog advantages of their own. Anthropic’s bet is less another multimodal demo and more a sales funnel of runnable blueprint + payment-network endorsement + consulting delivery: intelligence at the model layer, a one-week demo at the blueprint layer, compliance and rollout at the partner layer.
Compared with OpenAI’s narrative, Anthropic leans harder into merchants building their own agents while keeping the customer relationship—Visa’s line about merchants wanting more control over how AI meets customers is almost the thesis statement. That matches Anthropic’s broader enterprise posture of controllable deployment and configurable guardrails: not replacing the checkout provider, but owning the reasoning and orchestration layer before money moves. For platforms such as Shopify, Wix, and Square, the blueprint lowers scaffolding cost and can also soft-lock model choice into the Claude stack—partnership and channel competition at once.
Risks, limits, and disputes
Draw the factual boundary first. “35% larger carts” and “60% more likely to purchase” are vendor-relayed customer results without sample size, controls, or third-party reproduction. Accenture’s 85% / nearly-three-in-four trust figures are the same class of claim. Treating marketing testimonials as industry laws is the classic slide.
On product scope, payment is deliberately left outside the blueprint: an agent can build a cart, but settlement, chargebacks, and fraud models still sit with the merchant and networks such as Visa and Mastercard. Catalog-true pricing and “no manipulative upsell” are only as strong as the deployer’s refusal to re-enable gray promotional prompts after fork. Human approval before merchant-agent changes is the right default and a peak-season bottleneck; if teams disable the gate for speed, the risk profile flips overnight.
The commercial motive is plain: holidays, pipeline, co-branded partners. The live dispute is downstream: if nearly three in four respondents say they would trust an agent to buy for them, brand loyalty, price transparency, and dark-pattern enforcement will heat up in parallel. Agentic commerce amplifies conversion—and the blast radius of bad recommendations and discriminatory pricing.
Critic’s take
I read the September 2 post as Anthropic’s bid to sell an industry operating system: model quality is already “good enough,” and the scarce asset is a default architecture that spans catalog, inventory, payment trust, and human approval. Open-sourcing the blueprint lowers the demo tax; the partner network raises enterprise comfort. Together they move the fight from “who chats better” to “whose legal, risk, and engineering teams can nod in the same week.”
Focusing only on bigger carts undersells it. The sharper tell is that pricing and promotion suggestions default to human approval—Anthropic knows the trust crisis in agentic commerce is not copy tone, it is silent price changes and auto-launched campaigns. The real product fork ahead is who turns the gap between “suggested” and “live” into exportable compliance evidence, not just another chat transcript.
Outlook
Over the next six to twelve months, agentic commerce will move from pilot pages toward a default entry point in retail and travel apps. Foundation-model vendors will compete less on leaderboard screenshots and more on forkable blueprints co-signed by payment networks. Anthropic’s September delivery of August’s design essay is coherent timing. Watch three things next: whether 35%/60%-class outcomes survive more transparent methods; whether human-approval gates survive peak promo week; and whether regulators and platforms assign liability for agent-led purchases to the model provider, the merchant, or the payment network.
Factual claims follow the Claude Blog posts; partner testimonials and survey figures are attributed as such above.