Friday, November 13, 2026 was the day agentic commerce stopped being a conference track and became shipped infrastructure. Three moves landed within hours of each other. OpenAI launched ChatGPT Wallet, the first wallet that lets an AI agent pay with a real card while never seeing the card number, a launch we analyzed in depth on this blog two days ago. Sierra released version 1.0 of the Personal Agent Protocol, known as Poppy, and Shopify switched it on for every merchant on its platform while Meta made it the default for Muse. And Google opened a Gemini integration with Google Wallet for all US users that can read every transaction and never move a cent.
The wallet solved payment. What shipped alongside it completes a different picture: the permission layer, the thing that decides what an agent may do at a business before it ever reaches a checkout, is now live at a scale most standards never reach. And Google’s deliberately crippled integration shows the industry’s other school of thought about how much power an AI should hold at all.
Put the week together with what shipped before it and the whole agentic commerce stack is now assembled. Web Bot Auth verifies that a signed agent is allowed through the firewall. Poppy defines what the agent may do once inside. UCP fills the cart. ChatGPT Wallet settles the charge. Identity, access, action, payment: every layer of delegation now has a protocol, a vendor, and a rollout plan.
What no layer addresses is the question that comes before all of them: is the product the agent is about to buy actually any good? The stack verifies the agent, the permission, and the payment. It verifies nothing about the purchase itself.
What Sierra shipped: the rulebook, at retail scale
Sierra’s Personal Agent Protocol defines what happens when an agent arrives at a business: how it identifies itself, what actions the business permits, and how a customer controls that access over time. Version 1.0 shipped November 13, and the adoption numbers attached to it are the kind most standards never see in a decade:
- Shopify has enabled Poppy for every merchant on its platform, across more than 175 countries. Every Shopify store already supported the Universal Commerce Protocol for finding products and checking out; now each one automatically publishes a Poppy file, so merchants need no extra implementation work.
- Poppy is on for everyone who uses Muse, Meta’s personal agent, which “automatically looks for Poppy first on every site it visits.” Muse passed 10 million US downloads within 40 days of its September launch, a milestone only 11 other apps have ever reached, and shopping is already one of its top activities.
- More than 100 companies are now design partners, after 60 joined in this round alone: Affirm, Agoda, Checkout.com, Delta Air Lines, HelloFresh, Hyatt, Intuit, Lemonade, Navy Federal Credit Union, Salesforce, SoFi, Travelers, Twilio, Wayfair, Wix, and Zoom among them.
- Three partners are already live beyond the protocol itself. Mercury lets agents check balances and initiate payments that route through the account’s existing approval policies. Turkish Airlines lets agents pull fares, flight status, and even business-class upgrades that customers approve before paying. Synchrony is exploring financing offers surfaced inside agent-led shopping across a network of more than 500,000 merchant locations.
The division of labor between the two Shopify-backed protocols is the most instructive part. UCP handles the shopping: finding products, filling a cart, checking out. Poppy handles everything before that, the moment an agent shows up at a store, identifies itself, and learns what it is allowed to do. Shopify co-authored the commerce extension that links them, letting an agent shop as a guest or with the customer’s account, and forcing it to ask the customer before taking sensitive actions.
Note the engineering humility in the design. Poppy is not a new wire protocol; it is built on existing standards, including MCP, OpenAPI, and UCP, with OAuth for authentication. The permission layer did not invent a transport. It standardized the handshake.
Now note what the handshake contains. Identity, authorization, scope, duration. Nothing in it asks whether the recommended product deserves the purchase. That is not a flaw in Poppy; permissions were never going to carry evidence. But it means the most-deployed new standard in agentic commerce is a gate that checks the visitor, not the goods.
Google’s answer: the read-only agent
The same Friday, Google took the opposite door from OpenAI. Gemini’s Google Wallet integration, rolling out to all US users across Android, iOS, and web, lets the assistant read transaction history, chart cash flow, flag rising subscription charges, and build holiday budgets from Gmail and Wallet data, with optional deeper analysis through Plaid-linked accounts.
The permission boundary is absolute. Gemini can read, and can do nothing else: no payments, no transfers, no investments. The connection is off by default, toggled by the user in settings, severable at any time, and Google added a “Memory Summary” view so users can audit exactly what the AI remembers about them.
The banks have already picked sides in this design argument. Per market reporting on the two launches, Citigroup, Bank of America, and Capital One permit user-authorized AI access to their cards, while JPMorgan Chase explicitly prohibits AI access to its credit cards. When the largest US bank and the largest AI company disagree about whether an AI should ever hold payment credentials, the read-only position is not timidity. It is a product strategy with a constituency, and it maps onto a real distribution of consumer preference.
The honest read is that both designs will coexist, and together they reveal what the industry is actually confident about. OpenAI is confident it can secure the transaction. Google is confident it can be useful without it. What neither company claims, anywhere in either announcement, is that it can make the agent choose well.
The stack, assembled and incomplete
Line up everything that now exists and the layers are all shipped:
- Access: Web Bot Auth, promoted in OpenAI’s Commerce Partner Playbook, lets a site verify a signed agent at the firewall instead of blunt-blocking it.
- Permission: Poppy 1.0 tells the agent which actions a business permits, with OAuth-scoped customer control.
- Transaction: UCP finds products and fills the cart; ChatGPT Wallet, Visa’s Trusted Agent Protocol, Mastercard’s Verifiable Intent, and Google’s AP2 settle it.
- Payment credential: ChatGPT Wallet’s network-issued tokens let an agent pay without ever holding the card number.
Every layer answers a question about the agent or the transaction: who is this agent, may it act, may it pay, with whose money. That is the correct order of operations for infrastructure, and the engineering is genuinely strong.
Now ask the assembled stack a different question: should this purchase happen at all? Is that 4.8-star rating organic, does the best-seller badge reflect quality or ad spend, is the listing the agent surfaced the right product or merely the best-marketed one? Silence. The evidence gap stopped being theoretical the moment the wallet made it chargeable. Visa’s Agentic Commerce Benchmark, released November 5, put 22 models from 12 labs through realistic shopping runs: 42.8 percent of runs failed, only 34.2 percent of scenarios passed three consecutive attempts, and in 54 percent of failed runs the agent bought something anyway. Last month those failures were lab findings. This month the same failure mode connects to a tokenized Visa card through an approval flow the user trusts precisely because the payment layer is so well built.
Feeds are marketing: the quiet assumption in the new stack
There is a subtler problem hiding in the week’s announcements, and it deserves its own heading. Every path an agent now has to product information runs through the seller.
Poppy’s handshake tells the agent what it may do at a store. UCP then hands it the store’s own catalog. OpenAI’s playbook, for its part, asks merchants to claim their business and upload product feeds with pricing and availability through the Business Hub. Necessary and sensible steps, all of them. But merchant-supplied feeds are marketing assets, authored by the party with the strongest interest in the outcome. A stack where the only structured product data an agent consumes is seller-published is a stack where the agent inherits the seller’s incentives at machine speed.
This is not paranoia; it is the pattern already documented in the human version of the market. The FTC’s ongoing case over Amazon’s ad auction alleges that the surfaces consumers and tools treat as neutral, placement and ranking, were systematically reshaped by ad spend. Agents do not read those surfaces more skeptically than people do. They read them faster, at scale, and soon with a payment credential attached.
The permission layer governs whether the agent may buy. The payment layer governs how. Nothing governs whether the evidence justifying the purchase is worth anything.
What evidence-first delegation looks like
The fix is not to slow the protocols down. It is to give agents an evidence source independent of the sellers being evaluated, consulted before the buy-or-decline fork. That is the layer GoBuy builds, and this week’s standardization makes it easier to wire in than ever:
- Smart Score 0 to 100, computed from review quality, not review quantity. A thousand purchased five-star reviews score below two hundred verified, substantive ones, which breaks the legible-signal shortcut agents otherwise regress to under uncertainty.
- Fake review filtering before scoring. Incentivized, coordinated, and generated reviews are suppressed upstream, so the rating an agent reads is the rating that survived manipulation rather than the one that was purchased.
- Top 7 products per category, not thousands. A shortlist of defensible picks maps onto exactly what Visa’s benchmark rewarded: the models that uncovered the most customer requirements passed 70.2 percent of runs, because disciplined discovery beats confident guessing across a polluted shelf.
- GoBuy Verified. Products holding a Smart Score above 80 across a full 90 days earn a durable badge, giving an agent a stable pass-or-fail input rather than a snapshot that can be gamed for launch week.
All of it is served over MCP at gobuy.ai/api/mcp, and the timing is the point: Poppy itself is built on MCP, and OpenAI’s playbook explicitly invites merchants to expose WebMCP tools. The agentic stack has converged on the exact protocol a trust layer needs to sit inside the decision flow, before the wallet gets involved. The agent’s question sequence should be: what does the customer need, what does independent evidence say, then pay. The stack shipped this week answers the third question beautifully and still skips the second.
The week the money moved
A final framing on where this heads. DoorDash has reported only several hundred agent-placed orders, and Uber says almost no one books rides through agents, which skeptics cite as proof of weak demand. The better reading is sequencing: payment and permission rails had to exist before demand could express itself, and those rails arrived this month. Amazon inviting a subset of shoppers to text Alexa Shopping directly, United restricting unauthorized booking agents, JPMorgan prohibiting AI card access, and the Federal Reserve’s stability survey now flagging AI as the top-cited concern for 87 percent of contacts, up from 50 percent in May: these are the flanking motions of a transition that is irreversible in direction and undecided in design.
The companies that shipped this week solved authorization, and solved it well. The open prize is the layer underneath every authorization: knowing that what the agent buys is worth buying. Wallets will multiply, permission protocols will consolidate, and the first stack to make agents reliably right, not just reliably permitted, will own the trust position in agentic commerce.
If you are building a shopping agent, an agentic checkout flow, or adapting a storefront for Poppy and UCP, wire product evidence in before the payment rail: read the agent docs, connect to the MCP server at gobuy.ai/api/mcp, and give your agent a reason to decline before it earns the right to pay.