anthropic's commerce agent blueprint: what it gives you, what stays yours.
anthropic published the code for an agent that shops on a customer's behalf. it left out the part that takes the money, and that omission tells you exactly what your job is now.
anthropic released an open-source blueprint on september 2, 2026 for two commerce agents: a shopping agent that searches, compares and builds a cart, and a merchant agent that handles listings, inventory, pricing and campaigns. it ships under apache 2.0. it does not ship payments, a product catalog, authentication or ads. those stay yours.
what anthropic actually shipped
it is a reference implementation, not a product. the repository is anthropics/commerce-agents, licensed apache 2.0 and marked "Copyright 2026 Anthropic PBC". you get two agent definitions, ten skill flows between them, and four runnable reference verticals: retail, travel, telecom and entertainment.
the architectural choice is worth noting because it contradicts how most people build agents. anthropic uses one agent with modular skills rather than a swarm of subagents, and runs the same prompt, skills and tool contracts across the messages api, the claude agent sdk and managed agents. if you have been told more agents means better results, this is the counterexample from the people who make the model. it is the same skills-over-prompts pattern i wrote about in prompts vs skills.
accenture, mastercard and visa are working with anthropic to bring the blueprint to their clients and merchant networks (PYMNTS, 3 sep 2026).
what the blueprint gives you and what stays yours
this is the whole story in one table. anthropic built the reasoning layer and deliberately stopped at the edge of your business systems.
| piece | in the blueprint | still your job |
|---|---|---|
| agent reasoning | yes, two agents, ten skills | tuning to your catalog and tone |
| payments and checkout | no payment protocol at all | your processor, your rails |
| product catalog | no, sample data only | your product data, structured |
| authentication | none | your customer accounts |
| advertising layer | none | your paid channels |
| writes to live systems | staged for approval, never executed | who approves, and how fast |
| being findable by the agent | not covered | entirely yours |
that last row is the one almost nobody reads. the blueprint assumes the agent already has a catalog it can query. if an assistant cannot find or parse your store in the first place, none of the rest applies to you yet.
the two agents, and the ten skills between them
the split is clean: one agent faces your customer, one faces your staff. neither does the other's work. the skill names below are the directory names in the repository, not my labels.
the shopping agent "searches, compares, plans, fills the cart, answers order and policy questions, and remembers what a customer tells it." its five skills:
search-discovery, finding the right item in your catalogpurchase-research, comparing options before a decisionplanning-goals, building toward an outcome rather than a single itemcustomer-care, order and policy questionsmemory-personalization, remembering what the customer said
the merchant agent "explains performance, maintains listings, acts on inventory and order alerts, prices and promotes, and drafts campaigns." its five skills:
performance-insights, explaining what sold and whycatalog-listings, maintaining listingsinventory-operations, acting on inventory and order alertspricing-promotions, pricing and promotingmarketing-campaigns, drafting campaigns
read the merchant list again. that is a description of the back-office job most solo sellers do badly at 11pm, and it is the half of the blueprint nobody is writing about.
nothing in it can spend your money
the most important line in the repository is a limitation, not a feature. anthropic writes:
"Nothing places an order, charges a card, or changes a live listing"
every write is staged for a human to approve. no authentication, no live payments, no ad systems. anthropic built the shopping brain and skipped the wallet, so no transaction routes through anthropic (PYMNTS, 3 sep 2026).
anthropic reports that retailers running shopping agents on claude see carts up to 35% larger and shoppers 60% more likely to complete a purchase. those are anthropic's own reported figures rather than an independent benchmark, which is how PYMNTS characterised them on 3 sep 2026, and you should read them as a vendor claim until someone outside the company measures it.
why this matters if you are not a retailer
because the buying behaviour is already real, and it is measured by someone with no stake in anthropic. adobe analytics found that ai-referred traffic to us retail sites converted at a rate 60% higher than non-ai traffic in july 2026 (adobe analytics, released 19 aug 2026). that was the eleventh consecutive month ai-referred traffic outperformed other channels on conversion rate (adobe analytics, 19 aug 2026). ai-referred traffic to those sites also rose 62% year over year in july (adobe analytics, 19 aug 2026).
so the traffic converts better, the volume is climbing, and the tooling to act on it is now open source. the constraint has moved. it is no longer "can an agent shop for me". it is "can an agent read my store".
that is a different problem from search rankings, and i have written about the shift in the assistant is the new homepage and optimize for the assistant, not the browser. the mechanics of how a model chooses what to name are in how chatgpt decides what to recommend.
what to fix first if you sell anything online
the answer is boring and it is not a rebuild. make the facts an agent needs readable as text.
- put price, availability, variants, shipping cost and return window in plain text on the product page, not baked into an image or a pdf.
- write the policy answers out. an agent handling "can i return this after 40 days" needs a sentence it can quote, not a chat widget.
- check what assistants say about your product today. that output is the starting state an agent inherits, and it is often wrong in a way you can fix in an afternoon.
- keep a plain-text file describing what you sell and to whom. llms.txt is one convention for this, with real limits worth knowing.
the takeaway
anthropic gave away the hard part and kept none of the money. the blueprint is the reasoning layer, free, under apache 2.0, with payments and catalog and ads left deliberately in your hands. that means the differentiator is not whether you have an agent. it is whether your store can answer one.
common questions
does anthropic's commerce agent blueprint let an ai actually take payment from my customers?
no. the blueprint ships no payment protocol and no checkout processor, and no transaction routes through anthropic. the repository states that nothing in it places an order, charges a card, or changes a live listing: every write is staged for a human to approve. payments stay with your existing store and payment partners.
do i need to be a big retailer to use the claude commerce agent blueprint?
no, but you do need an engineering resource. it is source code released under apache 2.0, not an app you install. a solo seller on a hosted store cannot switch it on. the useful part for a small seller is the pattern it documents: what a shopping agent asks your store for, and whether your product data can answer.
what is the difference between the shopping agent and the merchant agent?
the shopping agent faces the customer. it searches, compares, plans, fills the cart, answers order and policy questions, and remembers what a customer tells it. the merchant agent faces your staff. it explains performance, maintains listings, acts on inventory and order alerts, prices and promotes, and drafts campaigns. they are two separate agents with five skills each.
if i sell online, what should i fix before an ai agent tries to buy from me?
make the answers machine-readable. an agent cannot read your price off an image, infer your return window from a pdf, or guess stock from a sold-out badge rendered in javascript. put price, availability, variants, shipping and returns in plain text on the page. then check whether assistants describe your product correctly today, because that is what the agent starts from.
find out what assistants actually say about you
the ai visibility scan checks whether models can find, read and correctly describe your business, and shows you the specific gaps to fix first. free, no call.
run the free scanbuilding this stuff alongside other people is easier than doing it alone. the ai builders lounge is where the weekly builds happen.
or just follow along. new field notes most weeks on x, instagram, and tiktok.