Your Agent Now Knows Your Catalog, Your Policies, and This Customer. Where You Put It Decides Whether It Works.

The Founders’ AItrained on 25 years of our work

Want this made concrete for your company? Ask me, or pick one:

Where should our first commerce agent actually sit? What should I do first? Take the 90-second readiness read

Ask

What an agent that knows your business can now do

An agent that has your catalog, your policies and this specific customer in front of it can hold a different kind of conversation than anything a filter UI has ever managed.

It can take a described need rather than a search term. A gift, a constraint, a project, a job to be done. It can narrow to a shortlist and explain why those and not the others. It can answer what your return window actually is, in your words, from your own policy document, at eleven at night. It can remember that this person buys wide fit, and it can tell them when the thing they wanted is available for pickup at the store they usually visit, at the top of the hour, inside opening times.

On the staff side it can read your sell-through, tell an owner which segment moved a number and why, notice that a line is running short before anyone opens a spreadsheet, and draft the promotion. Then it waits for a person to approve it.

None of that is speculative. On September 2, 2026, Anthropic published working code for both halves of it. Claude Commerce Agents is an Apache 2.0 repository at github.com/anthropics/commerce-agents containing a customer-facing shopping agent and a staff-facing merchant agent, with runnable implementations for four verticals: retail, travel, telecom and entertainment. Cloned and counted, it is 213 Python files, 41,445 lines of Python and 83 test files.

It is also, by its own README, a blueprint rather than a product. The disclaimer is one sentence: “This is a reference implementation; it is not maintained and does not accept contributions.” Every company in the demo data is fictional. Whoever deploys it owns that fork permanently.

Two things get called an agent in a chat window, and the industry uses one phrase for both. One is a conversation running on the merchant’s own site, powered by somebody’s model. The other is a conversation inside a general assistant the merchant does not own. Both look identical to a shopper. They are opposite businesses. One compounds a relationship you already have; the other rents you an audience on terms set by someone else. Most disagreements about agentic commerce are two people using one sentence about two different places.

The release is built for the merchant’s own site, and its value there has little to do with running the code. It is the most precisely articulated public description of how a commerce agent should be built, and it is readable by anyone, on any stack. You can study the architecture and implement it on Microsoft’s tooling or Google’s.

The same technology, two placements, two results

Until recently the assumption across the industry was that the shopping agent would live in the assistant. You would ask ChatGPT for running shoes and buy them there, and the merchant’s website would become a supplier to a conversation happening somewhere else.

That thesis got a real test. On September 29, 2025, OpenAI and Stripe released the Agentic Commerce Protocol and launched Instant Checkout inside ChatGPT, letting shoppers complete a purchase without leaving the conversation. On March 24, 2026, just under six months later, OpenAI stepped back from it, returning checkout to the merchants and moving its own effort to product discovery. In its words, the first version “did not offer the level of flexibility that we aspire to provide, so we’re allowing merchants to use their own checkout experiences while we focus our efforts on product discovery.”

The number underneath that decision came from the flagship partner. Daniel Danker, Walmart’s EVP of AI acceleration, product and design, told WIRED that those in-chat purchases “converted at one-third the rate of click-out transactions,” and called the experience unsatisfying.

None of that is a verdict on agents. Walmart’s catalog did not change. The model did not get worse. On Danker’s own figure, the same shopper, wanting the same thing, converted three times better one click away from the conversation than inside it. The variable that moved was where the buying happened.

The technology was constant. The placement was not.

That comparison is one retailer, one product and six months, which makes it the best evidence available and not a law of physics. But it is also the only like-for-like comparison anyone has published, because it compared two placements of the same catalog for the same retailer, which is more than any vendor number on either side of this argument can say.

Anthropic is unambiguous about which of the two it built for: “The shopping agent lives inside your app or website.” The solutions page is more pointed still, describing shoppers who get help “finding, comparing, and deciding inside your app, then check out without going to a marketplace or third-party assistant.” The blueprint “leaves payment to you, whether that is your existing checkout or an agentic payments provider.”

That is enforced in the code, not merely asserted in the marketing. The shopping agent’s own system prompt instructs the model that “checkout stages a summary the customer confirms in the app; it places no order and charges nothing, and your text must not suggest otherwise.” The checkout URL, when the host supplies one, “is never a tool argument and never reaches the model.” There is no payment integration anywhere in the repository.

Asked directly whether it wants to be a shopping destination, Anthropic’s answer on its own site is one word and then a boundary: “No. Claude is your intelligence layer, not a storefront or checkout. We’re not interested in owning catalogs, supply chain, or the last mile, and there’s no advertising or paid placement of Claude in any agent you build.”

Anthropic also reports results, and they should be read with the label on. Its launch material states that “retailers running shopping agents on Claude have seen carts up to 35% larger and shoppers 60% more likely to complete a purchase.” No client is named against those two figures, no sample size is given, no methodology and no baseline. A separate claim of 40 percent higher conversion for small businesses does carry a company attribution, to Vambe, and still no sample or method.

Put those beside Walmart’s third and the lesson lands somewhere other than either number. Agentic commerce results move with placement, and nobody on the merchant-property side of the argument has yet published a clean measurement. Both of those statements are true at the same time, and a retailer making a decision this quarter deserves to hear both.

What kind of commerce are you, actually

Before choosing a platform, work out which problem you have. The answer is not your industry label.

The blueprint ships four runnable implementations, and reading them side by side makes something clear that no vendor page says out loud. They are not four industries. They are four shapes of inventory, and the shape is what decides how hard your agent is to build and where it can fail.

Countable units, no time dimension. This is the retail baseline. A thing exists or it does not, you have eleven of them, and eleven tomorrow is the same as eleven today. Everything else in this section is retail plus a complication.

Date-bound capacity. The travel implementation exists because a search here cannot return a fixed price. It returns a quote for a context: these dates, this occupancy, this length of stay. The repository is explicit that a nightly rate for searched dates is not a product variant, and that if you do not declare rate fields as price-bearing, your own price guardrails will not check them. If your inventory is capacity sold against a calendar, your agent is closer to a travel agent than a shop assistant, whatever your industry is called.

Regulated disclosure over an installed base. The telecom implementation reads account context on every turn, serves fee disclosures written by the server instead of by the model, and marks regulated fees as protected fields the agent cannot touch. A price change here is not a catalog edit. It reaches live subscriptions, so the system makes a price move declare how many lines it affects before anyone can approve it.

Perishable inventory with a clock. The ticketing implementation holds seats on a server timer, runs waitlists, and computes what is left as capacity minus sold, minus held, minus offered. Nothing here survives the event date. If your inventory expires, the hard part is not recommendation, it is arithmetic under contention.

The reason this matters commercially is that it reassigns businesses that assumed they were out of scope.

Take a moving company. No catalog, no cart, no checkout, so the retail shopping agent is irrelevant and most operators stop reading there. But a move is date-bound capacity priced against a survey, which is structurally the travel problem. The object is the quote rather than the cart. A reference implementation for a moving quote agent already exists in that repository, and nobody has called it that, because it is filed under travel.

The same reassignment applies to a specialist manufacturer selling configurable, spec-driven products, where the buyer assembles a compliant basket against requirements rather than browsing. That is not retail either. It sits closer to the telecom pattern, where disclosure and eligibility do more work than recommendation.

And it explains a failure that gets blamed on the wrong thing. When a conversational agent underperforms on a business with date-bound or perishable inventory, the usual diagnosis is that the model was not good enough. The more common cause is that the thing was built on the retail pattern, which quietly assumes an item’s availability and price do not depend on when you ask.

So the first question is not which vendor. It is which of those four shapes describes what you sell, because that determines what your agent has to be right about, and being right about the wrong thing is the expensive failure.

Layer one: the agent on your own property

Start with one agent, on a surface you control, built so the model underneath it can be swapped later.

One agent, not a fleet. Anthropic’s companion guidance is unusually clear on this and the reasoning generalizes: a commerce conversation is one tightly coupled session that moves between discovery, comparison, policy and after-sales without warning, and every handoff between specialized agents loses context and adds latency. So the pattern is a single agent with skills, which is the same conclusion we reached writing about why AI skills beat AI agents as an architecture.

Which platform depends almost entirely on what you already run. This is an ordinary stack decision that people keep dressing up as a strategy decision.

Two things are worth carrying into that conversation.

The first is a correction. You will hear that only one vendor has published commerce-specific agent material, and that is wrong. Microsoft shipped retail agent templates in January 2026 and Google sells a managed shopping-agent product. What is true, and narrower, is that no one else has published an open, permissively licensed, end-to-end reference codebase covering both a customer-facing and a staff-facing agent, with a safety harness, four runnable verticals and a published evaluation method. That is a difference in how much design work has been done in public, not a difference in model capability, and the design is portable to any of them.

The second is about lock-in, and it has a date attached. OpenAI launched Agent Builder at DevDay on October 6, 2025, deprecated it on June 3, 2026, and will shut it down on November 30, 2026. Under eight months from launch to end-of-life notice. Take that as an argument for keeping the model and the runtime replaceable, which is exactly the discipline we laid out in what a model switch actually costs you. The blueprint’s own structure separates those concerns already: backend interfaces, skills and safety harness on one side, the model and the prompts on the other.

Layer two: presence inside agents you do not control

The second layer is different in kind, and conflating the two is the expensive mistake.

Here you do not pick one. You need to be present everywhere, and it takes one body of work: feeds, protocols and schema.

Syndication does most of the work. Shopify’s Agentic Storefronts pushes a merchant’s catalog into ChatGPT, Google AI Mode and Gemini, Microsoft Copilot and Meta, per Shopify’s own current channel list. One integration, several destinations. Checkout behaves differently by channel: for ChatGPT, shoppers complete the purchase on your own store checkout inside an in-app browser, while on other channels a merchant who has activated direct checkout lets customers finish inside the channel. Shopify hands you that choice, and the Walmart datapoint is the reason to think about which way you set it.

Underneath syndication sit the protocols. The Agentic Commerce Protocol from OpenAI and Stripe. Google’s Universal Commerce Protocol, announced by Sundar Pichai at the NRF keynote on January 11, 2026, developed with Shopify, Etsy, Wayfair, Target and Walmart, and compatible with A2A, AP2 and MCP. Then the unglamorous layer that decides whether any of it works: clean product schema, an llms.txt file, a stated AI-crawler policy, and content an agent can read without executing JavaScript. That is the same architecture argument we made in how agentic protocols are replacing discovery, evaluation, and transaction, and it has not changed, it has only got more urgent.

Layer two is cheap, model-neutral and fast. For most mid-market retailers it is where to start.

The protocol half of it moves fast and lands slowly, and both halves of that sentence matter. Google donated AP2 to the FIDO Alliance on April 28, 2026 and shipped a version introducing “Human Not Present” payments, which “will allow agents to securely execute payments autonomously, like securing and purchasing limited-run tickets the moment they’re on sale, based on pre-authorized user instructions.” That is a standards body building directly toward the thing this article says has not worked yet, with Mastercard, Visa, American Express, PayPal, Stripe, Adyen and Microsoft alongside it.

Set against that, the one in-chat checkout that has actually shipped uses none of these protocols. When we mapped this stack in June we treated the protocol layer as the mechanism by which agentic commerce would arrive. Six months on, the mechanism that arrived first was something else entirely.

That layer is also changing shape.

Until recently, being present inside an assistant you do not own meant being findable there. Your catalog got syndicated, an agent read it, and if the shopper wanted to buy, they left. That is still how most of it works. Claude’s own connector directory is full of it. Glovo’s entry says so in one sentence: “When you’re ready to order, Claude takes you directly to the Glovo app or website to complete your purchase.” Uber, Uber Eats, Booking.com, StubHub and Instacart all do a version of the same thing. Search, compare, hand off.

Then in January 2026, Anthropic shipped an extension letting third parties render their own interactive interface inside a Claude conversation. Three months later a payments company used it. Block’s Order by Cash App connector, listed in Anthropic’s directory in April 2026 and announced by Square that July, lets a shopper “discover nearby restaurants, compare menus, customize your order, and check out, all in the conversation.” The connector exposes no checkout tool and no payment tool. What it exposes is an ordering view, drawn by Square, inside somebody else’s chat window.

That is a different thing from a marketplace, and the distinction is the whole point. Anthropic did not build a checkout. Where one now exists inside Claude, a payments company brought its own, kept the payment on its own infrastructure, and pays no commission for the placement. The exception is narrow today: United States restaurants, sellers with a Square ordering profile already active, one connector out of nearly eight hundred.

But the direction is legible. The assistant is becoming a surface you can bring your own transaction to, on terms that differ per platform, rather than a single gate you either get through or do not. Which means Layer two is no longer only a discoverability question. Increasingly it is the same question as Layer one, asked about a surface you do not own: what are you willing to put in there, who holds the customer relationship afterward, and what happens to it when the platform changes its terms.

One version of this belongs to a business with no ecommerce at all. If your inventory sits in stores, the equivalent mechanism is Google’s Local Inventory app, the service formerly called Pointy. It pushes in-store stock from your point-of-sale system into Google Search, Maps and the “is it in stock near me” answers that AI Mode began rolling out in the United States in April 2026.

The bar is low. Google’s specification requires four attributes per item: an id, a store code, availability, and a pickup service level for anything available for in-store collection. It is free, it is organic, it works with more than twenty point-of-sale systems including Square, Clover and Lightspeed, and it is live across Europe. Every other mechanism in this section assumes an ecommerce catalog exists first. This one does not, which makes it the cheapest way a physical retailer can appear in an answer they will never hear being asked.

The agent most of you should build first is the one nobody is selling

Everything above concerns the customer. The more interesting half of the September release faces the other way, and almost nobody is writing about it.

The merchant agent is a staff tool. It explains why a metric moved and which segment drove it. It maintains listing content. It watches stock and capacity and raises the low-stock and slow-mover alerts. It proposes price changes inside your guardrails. It drafts campaigns. It produces the operator’s morning digest.

What makes it the honest first move is not the feature list. It is the failure mode.

If a customer-facing agent quotes the wrong price or invents a return policy at two in the morning on Black Friday, that is a customer, a refund and possibly a regulator. If a staff-facing agent proposes a bad price change, a person looks at it and rejects it. Those are different classes of risk, and the second one is survivable in a way the first is not.

The architecture is what makes that true, and it is enforced in code rather than requested in a prompt. In the reference implementation, of nineteen merchant methods exactly one writes to a live system. Five stage proposals that touch nothing. The rest only read. Anthropic’s documentation states it directly: “Reads are free to call; stage_* methods record a proposal without touching live state; only apply_change mutates anything.” A price move has to declare how many lines it affects. Regulated fee fields are marked protected and cannot be modified at all.

The line that should end every debate about whether an approval gate is real: with host approval switched on, a change applies only for ids your own code has marked approved, because “a preview card approves nothing; an approval typed in chat sets nothing.”

This is the Humans First argument in its most concrete available form. Not a value statement about keeping people in the loop, but a system in which the machine is architecturally incapable of acting alone, and the human authorization is a fact in the code rather than a promise in a policy. It is the same principle we described as progressive trust in Anthropic’s safety framework for AI agents, with the gate moved from the model into the harness.

And the pattern is not one vendor’s. Square shipped Managerbot into open beta on April 28, 2026, across most non-franchise United States food and beverage, retail, and health and beauty sellers, at no additional cost. It watches sales and labor, flags inventory shortages, drafts schedules and campaigns. Square’s description of how it works could have been written by the same architect: “Managerbot proposes an action, and sellers approve. Every action requires the seller’s approval before it executes.” Different company, different model stack, identical shape. When two competitors independently land on propose-then-approve, that is a convergent design, not a marketing position.

Two of the largest physical retailers in the world arrived at a version of the same conclusion. Lowe’s put Mylow Companion in front of associates across more than 1,700 stores. The Home Depot expanded Magic Apron on August 27, 2026 into all 2,000-plus of its United States stores, serving customers and associates together, with Jordan Broggi, EVP of Interconnected Retail, describing it as taking “the best of our digital expertise” and bringing “it into the physical store.”

For a mid-market retailer the practical read is this. The staff-facing agent carries no conversion risk and no brand exposure, it runs against systems you already own, and you can ship the first version with every write path switched off. Anthropic’s own definition of a starting deployment is a merchant pilot that implements the eight read methods and lets the writes refuse. Six of those eight are implementable from a point-of-sale system alone.

Which answers a question we get from retailers who run stores and very little ecommerce: no, an agent does not mean exposing your point-of-sale system to your customers. It means letting an internal tool read from it, and deciding separately, later, whether anything is ever allowed to write back.

What the evidence actually supports, and what it does not

A piece that argues placement decides outcomes owes you the strongest objection to itself, and the strongest one cuts against the merchant-property side rather than for it.

Forrester’s version is the mild one, from March 2026: “Agentic commerce is not the ‘death of the retail website.’ Websites will need to change, but they’re not over.” Fair, and already conceded above.

The sharper objection is about measurement. Every published number on merchant-hosted agents compares people who chose to use the assistant against people who did not. Macy’s reported revenue per visit roughly 4.75 times higher among Ask Macy’s users during its beta. Lowe’s CEO Marvin Ellison told the August 19, 2026 earnings call that online customers using Mylow are “three times as likely to convert as customers who do not use the tool.” Neither is a holdout test. Shoppers who open a styling assistant are shoppers who already intended to buy.

The practitioners are saying so out loud. In an April 2026 Retail Systems examination of whether AI shopping assistants really lift conversion, Julian Skelly of Publicis Sapient put it precisely: “A 25 per cent uplift in users who engage with the AI assistant differs from uplift across the full digital estate.” James Heimers of RAPP added that “in most cases this is a relative uplift on a low baseline.”

So the honest position is this. Nobody has published a controlled measurement of a merchant-hosted shopping agent. Walmart’s is the only like-for-like placement comparison in the market, and it happens to compare two placements of the same catalog for the same retailer, which is why it carries weight the vendor numbers do not. Decide on architecture and control, and treat every conversion multiple you are shown, from any vendor including this one’s sources, as a claim about engaged users until someone shows you the holdout.

Consumer readiness is its own constraint. PYMNTS Intelligence, working with Worldpay, reports that 45 percent of consumers would be comfortable letting AI agents complete purchases on their behalf while 95 percent have at least one concern about agentic commerce. Neither figure comes with a published sample size or field dates, so treat them as direction rather than measurement. The shape is still informative: appetite and trust are not the same curve, and the gap between them is where brand damage lives.

Where to start, in order

If you do one thing, publish your product data so an agent can read it. Schema, a clean feed, and if you have stores, in-store stock into your Business Profile. Publishing it costs almost nothing, stays model-neutral, works whether or not you ever build an agent, and is the prerequisite for every other option on this page.

Then build the staff-facing agent, reads only, writes refusing. You will learn what your data is actually like, which is the finding most retailers skip past and then discover expensively six months into a customer-facing build.

Then, and only then, ask whether a customer-facing agent belongs on your property. By that point you will know whether your catalog and policy data can support one, which is the real gate. Most pilots underperform on data quality long before they underperform on model quality.

One part of this is genuinely hard to do from the inside. The engineering is the easy half. The hard part is deciding, against a vendor roadmap and an internal advocate who is already excited, whether the honest answer for your company this quarter is a staff-facing pilot, a feed, or nothing at all.

That decision comes easier to someone with no product to sell you, and it is the substance of the AI strategy and coaching work we do at bosio.digital. Whoever you ask, apply the same test to them that you would apply to us. At the end of the engagement, do you own the decision and the reasoning behind it, or do you own a dependency on the firm that made it? If a recommendation cannot survive being written down and handed to your own team, it was not a recommendation.

The architecture question was never which model. The models will change, and the good news is that none of the work above is wasted when they do. Your catalog will still need to be readable. Your policies will still need to be written down somewhere a machine can find them. Your stock counts will still need to be true. Every one of those is a thing you should have done anyway, and every one of them is now worth more than it was a year ago.

The question worth sitting with is the one your industry page has been asking all along. When every retailer can put an agent in front of a customer, what still makes yours worth choosing? The answer will not be the agent. It will be whatever the agent is standing on.

Sources

• Building Commerce Agents with Claude (Anthropic, 2 September 2026): launch announcement, the four verticals, and the reported cart and conversion figures
• anthropics/commerce-agents (GitHub, Apache 2.0): the reference implementation, the README disclaimer, the backend interfaces, and docs/safety.md
• Build commerce agents with Claude (Anthropic): named partners, the Priceline and Vambe references
• Powering product discovery in ChatGPT (TechCrunch, 24 March 2026): OpenAI handing checkout back to merchants, quoting OpenAI’s own post
• Walmart says ChatGPT checkout converted worse (Search Engine Land, 19 March 2026): Daniel Danker’s one-third figure, originally reported by WIRED
• OpenAI deprecations (OpenAI): Agent Builder deprecated 3 June 2026, shutdown 30 November 2026
• Stripe and OpenAI launch the Agentic Commerce Protocol (Stripe, 29 September 2025): ACP release and co-development
• Under the hood: Universal Commerce Protocol (Google Developers): UCP collaborators and protocol compatibility
• Agentic storefronts (Shopify Help Center): the current channel list and the checkout behavior by channel
• Microsoft propels retail forward with agentic AI capabilities (Microsoft, 8 January 2026): the shopping and store operations agent templates
• How Macy’s built the Ask Macy’s AI agent in 4 weeks (Google Cloud, 22 April 2026): Gemini Enterprise for Customer Experience and the revenue-per-visit figure
• Managerbot open beta (Square, 28 April 2026): the propose-then-approve architecture in a point-of-sale product
• Magic Apron expands in-store (The Home Depot, 27 August 2026): rollout across 2,000-plus stores, serving customers and associates
• Lowe’s AI assistant turns questions into sales (PYMNTS, 27 August 2026): Marvin Ellison’s earnings-call figure and the Mylow Companion rollout
• Do AI shopping assistants really lift conversion? (Retail Systems, 21 April 2026): the Skelly and Heimers critiques of engaged-user measurement
• What it means that the leader in agentic commerce just pulled back (Forrester, 7 March 2026): the counter-argument on the retail website
• Agents of Change: How Agentic AI Is Redefining Commerce (PYMNTS Intelligence with Worldpay): the 45 percent and 95 percent consumer figures, published without sample size or field dates
• About the Local Inventory app and local inventory data specification (Google): the former Pointy, the supported point-of-sale systems, and the four required attributes
• Transparency obligations under Article 50 of the AI Act (European Commission): applicable from 2 August 2026

Frequently Asked Questions