Georgia State University — J. Mack Robinson College of Business PATH — Pathways for AI Training & Hiring CIS 4394 Agentic AI  ·  Fall 2026  ·  Dr. Xinyu Fu
01 · Agentic commerce

Agents that buy.

Search is reversible. Summarizing is reversible. Paying is not. This page is about the four things that change when the last step of the loop is a purchase (irreversibility, liability, identity, and the merchant's side of the problem), and about the young, competing standards trying to make agent-initiated payments checkable.

Four shifts

What a purchase changes

1 · Irreversibility

Your retry logic was built on an assumption that just broke: that doing the thing twice is harmless. It is not. A timeout on place_order() is genuinely ambiguous: did the order go through? Retrying blindly is how one purchase becomes two. The fix has a name (idempotency) and it lives on page 02.

2 · Liability

When an agent buys the wrong thing, somebody eats the cost: the customer, the merchant, the agent's operator, or the payment provider. That allocation is decided by contracts and by what can be proven, which is why “what did the user authorize?” must be a question the system can answer, not a question the lawyers argue about afterward.

3 · Identity & authorization

Two different questions, often confused.

  • Identity: is this really the customer's agent?
  • Authorization: is this specific purchase inside what the customer allowed: this item, this amount, this window? A credential proves the first. Only an explicit record of the user's instruction proves the second.
4 · The merchant's problem

From the store's side, a helpful agent and a scraping bot look similar: automated traffic, no human rhythm, no mouse movement. Years of fraud tooling was tuned to block exactly that pattern. Merchants now need a way to tell “a customer's agent, acting with permission” apart from abuse, and to keep the sale rather than refuse it.

Notice the direction of travel. Every one of these four problems is solved by making something explicit that a human click used to leave implicit: what was authorized, by whom, for how much, and until when.
Interactive · click every step

The purchase flow, and where it breaks

This is the shape of nearly every agentic purchase. Click a node: each one tells you what goes wrong there and who ends up holding the cost.

rejected or edited 1 · intent + limits 2 · search / discover 3 · compare + build cart 4 · HUMAN APPROVAL 5 · authorized payment 6 · confirmation 7 · receipt + audit trail

The amber node is the last cheap place to be wrong. Everything above it can be redone for free; everything below it costs money to undo.

Start

Click a step in the flow

Work down the diagram. For each step, ask the two questions this page keeps repeating: what can go wrong here, and who pays when it does.

The design rule: push every irreversible consequence below the gate, and every correctable mistake above it. An agent architecture is largely a question of where you draw that line.
Emerging standards · read this carefully

Three young protocols, 2025 to 2026

In September 2025 two overlapping proposals appeared for making agent-initiated payments work: AP2, announced by Google with a set of payments and technology partners, and ACP, published by Stripe and OpenAI. In January 2026 Google added UCP, co-developed with Shopify, Etsy, Wayfair, Target and Walmart, which covers the whole shopping conversation rather than the payment alone. All three are early, all three are still changing, and they solve related but different problems. Treat what follows as a snapshot (checked November 2, 2026), not a settled answer.

AP2: Agent Payments Protocol

Making the authorization provable

Google's announcement frames AP2 as an open protocol for payments initiated by agents, built around mandates: signed, verifiable records of what the human authorized (the intent, the cart, and the payment), expressed as W3C Verifiable Credentials. The point is that a merchant or an issuer downstream can check the permission slip rather than take the agent's word for it, and that the record survives afterward as evidence of consent. It is positioned to sit alongside agent-to-agent and tool protocols rather than replace them (Google, 2025).

Primary source — the AP2 announcement ↗

ACP: Agentic Commerce Protocol

Making the checkout work at all

Stripe and OpenAI published ACP as an open standard for agent-initiated checkout: a documented way for an agent to hand a merchant a cart, have the merchant confirm price and availability, and complete payment, with a Shared Payment Token arrangement so the merchant continues to be the merchant of record and keeps its own relationship with the customer, rather than the agent platform taking it over. The specification is published openly (Apache-2.0 on the spec repository) so that merchants and agent builders can implement against the same shapes (OpenAI & Stripe, 2025).

Primary source — the Stripe announcement ↗ · agenticcommerce.dev ↗ · spec repo ↗

UCP: Universal Commerce Protocol

Making the whole shopping conversation portable

Announced by Google at NRF on January 11, 2026 and co-developed with Shopify, Etsy, Wayfair, Target and Walmart, UCP is an open standard for how an agent and a merchant talk from discovery through checkout, identity linking, order status and payment-token exchange. Its messages travel over REST, MCP or A2A, it is designed to work alongside AP2 for the payment leg, and its first home is checkout inside Google’s AI Mode and the Gemini app, paid with Google Pay. Endorsements at launch included Visa, Mastercard, American Express, Stripe and Adyen (Google, 2026).

Primary source — Google’s announcement, Jan 11, 2026 ↗ · under the hood (Google for Developers) ↗ · ucp.dev ↗ · spec repo ↗

AP2 (Google + partners, Sept 2025)ACP (Stripe + OpenAI, Sept 2025)UCP (Google + retailers, Jan 2026)
Centre of gravityPayment authorization: proving a human permitted this spend.The checkout transaction: a common way for an agent and a merchant to complete a sale.The shopping session end to end: discovery, cart, checkout, order status and identity linking, with the payment leg delegated to AP2 or a network token.
How consent is carriedMandates as verifiable credentials (intent → cart → payment), signed and checkable after the fact.The buyer's confirmed cart plus a shared payment token scoped to that purchase.The merchant confirms each step of the checkout; payment arrives as a token the merchant's processor accepts, or as an AP2 mandate.
Who most needs to careIssuers, networks, wallets, and anyone who must answer “was this authorized?”Merchants and agent platforms wiring up a working buy button inside a conversation.Retailers and platforms that want to be found and bought from inside AI assistants, starting with Google's.
OpennessPublished openly by Google with partner input; check the current specification before building.Specification published on an open repository under Apache-2.0; check the repo for the current version.Open specification at ucp.dev with an Apache-2.0 repository; co-developed by five retailers and endorsed by payment networks and processors at launch.
Maturity (as of this course)All three are early. They were announced between September 2025 and January 2026, all are still being revised, and they overlap in places and complement each other in others (UCP names AP2 as one of its payment options). Nothing here says one has won, and any page that tells you which one won is guessing.
How to talk about this in an interview (or a memo). Say what each protocol tries to make checkable, say that both are new and moving, and say what you would do about it: build the capability (an explicit record of authorization, a scoped payment credential, a receipt that links back to both) behind an interface you can re-point. Betting your architecture on one 2025 or 2026 standard is how you inherit a migration. Never quote adoption figures you have not read in the primary source this week; in a fast-moving standards race those numbers go stale faster than anything else on the page.
Special topic · checked November 2, 2026

How does the agent pay, and will the store let it in?

Meta’s Muse, released on September 8, 2026, is the first mass-market assistant whose job description includes buying things. Follow one Muse purchase from the request to the receipt and you meet every question on this page in order: whose card is this, who approved the spend, and whether the store wanted the agent there at all. Keep one more question running underneath: one day you may manage a company’s website. Will you let agents operate on it?

One Muse purchase, step by step

  1. You ask, in the Muse app or in WhatsApp. Muse runs on its own virtual machine with its own browser, so it can open retail sites, compare, and prepare a checkout while you do something else (Meta, 2026).
  2. It stops before it spends. Meta’s announcement says Muse comes back for approval before it sends an email or makes a purchase. CNBC describes the moment: Muse preps the checkout and shows an approval card with the order details for you to confirm (CNBC, Oct 3, 2026). That card is the amber node in the flow above, shipped as a consumer product.
  3. It pays with a card you never typed. Checkout runs through Link, Stripe’s wallet. In Meta’s words, “Link’s wallet for agents generates a one-time-use card so your real card details stay hidden.” Muse cannot see the card numbers stored in your Stripe account (PCWorld), and Meta says it is the first agent covered by Link’s purchase protections. Shop Pay and PayPal were announced as further ways to pay.
  4. The store decides whether to serve it. Walmart, Gap, Sephora, Expedia, Wayfair and Best Buy were named as partners in October 2026. Amazon blocked Muse, saying it violates Amazon’s terms of service and alleging that it stores customer credentials and scrapes account data (CNBC, Oct 3, 2026).

A code inspection reported by ppc.land on September 9, 2026 found three checkout paths inside the Muse client: a browser mode that types card details into whatever form the merchant presents, a Shopify mode over APIs (assessed, with a hedge, as Shop Pay over the Universal Commerce Protocol), and the Stripe Link path. Three paths, three different answers to “whose card, and who can check the approval?”

Sources: Meta newsroom, Sept 8, 2026 ↗ · CNBC, Oct 3, 2026 ↗ · PCWorld, Sept 25, 2026 ↗ · ppc.land, Sept 9, 2026 ↗

Four ways an agent can pay

Strip away the brand names and there are four designs. They differ in one thing: how far the credential has moved from “your card, in the agent’s hands” toward “a token for this one purchase that a third party can check.”

1 · The agent uses your card

A browser agent in form-filling mode (Muse’s browser path, Comet in the Amazon case) shops inside your logged-in account or types the card you stored. To the store it looks like you. Nothing scopes the credential to this purchase, and a planted instruction on the page (Week 10) is now spending your card. Every incident story on page 04 starts here.

2 · A wallet the agent never sees

Link by Stripe, PayPal, Shop Pay. The wallet holds the real card and issues a single-use number for the one order the human approved; the agent passes the token along without ever seeing a card. Purchase protection and the approval record sit with the wallet, which is why Meta chose this as Muse’s default path.

3 · Network tokens built for agents

Visa Intelligent Commerce and Mastercard Agent Pay (both announced in April 2025) extend the tokenization that already powers mobile wallets: a tokenized credential issued to a specific agent, with limits the cardholder set, so the issuer and the network know an agent is paying and can check the limits at authorization time rather than after the dispute.

4 · Protocol-level checkout

ACP’s Shared Payment Token (a credential scoped to one agent-initiated purchase, with the merchant staying merchant of record), AP2’s signed mandates (intent, cart, payment), UCP’s checkout with payment-token exchange. The approval travels with the transaction in a form the merchant, the processor and the issuer can each verify.

The pattern: each step down the list moves the human gate out of the agent and into a party that keeps records and absorbs disputes (a wallet, a network, a protocol). Where the gate sits decides who eats a charge nobody meant to make. It is the flow diagram’s rule, push irreversibility below the gate, applied to the credential itself.

Sources: Mastercard Agent Pay announcement, Apr 2025 ↗ · Visa Intelligent Commerce, developer page ↗ · Stripe, agentic commerce documentation ↗

How does the store know it is talking to an agent?

Amazon’s complaint against Perplexity turned on one word: disguise. A browser agent that presents itself as Chrome gives the merchant no way to treat it differently, so the merchant’s only options are to serve it as a human or to block everything that looks automated. Two 2025 mechanisms try to give the store a third option.

Sources: Cloudflare blog posts on Web Bot Auth ↗, signed agents ↗ and pay per crawl ↗ · Cloudflare press release, Oct 2025 ↗ · Visa Trusted Agent Protocol on GitHub ↗

Two postures a store can take

Block, and litigate: Amazon v. Perplexity
  • Nov 4, 2025Amazon sues Perplexity in federal court in California, days after a cease-and-desist accusing it of “disguising Comet as a Google Chrome browser” and refusing to identify its agents in the Amazon store. Amazon’s own agents, Rufus and Buy For Me, keep running; its CEO says Amazon expects “over time to partner with third-party agents.”
  • Mar 10, 2026Judge Maxine Chesney grants a preliminary injunction: Comet may not reach password-protected parts of Amazon on shoppers’ behalf. She found Amazon likely to succeed on its Computer Fraud and Abuse Act claims.
  • Aug 4, 2026A Ninth Circuit panel vacates the injunction. Judge Milan Smith: “An injunction would impair consumer choice and needlessly limit development of a nascent technology,” and it is the user, not Perplexity, who accesses Amazon’s computers. The case itself continues.
  • Oct 2026Amazon blocks Muse, citing its terms of service (CNBC).

Sources: Payments Dive ↗ · GeekWire ↗ · MediaPost ↗ · Engadget ↗

Join a platform, on a protocol
  • Sept 29, 2025Instant Checkout in ChatGPT opens with US Etsy sellers, with “over a million Shopify merchants” announced as coming, on ACP. The merchant stays merchant of record and decides whether to accept each transaction (Stripe).
  • Jan 11, 2026Google announces UCP at NRF with Shopify, Etsy, Wayfair, Target and Walmart as co-developers: checkout inside AI Mode and the Gemini app, paid with Google Pay.
  • Oct 2026Walmart, Gap, Sephora, Expedia, Wayfair and Best Buy are named as Muse partners.
  • The tradeWhat the retailer gets: the agent’s traffic arrives identified, the approval and the payment credential are checkable, and it keeps the customer record. What it gives up: the agent platform now sits between it and the shopper, and ranks it.

Sources: Stripe, Sept 29, 2025 ↗ · Google, Jan 11, 2026 ↗ · CNBC, Oct 3, 2026 ↗

👥 Group task · Small groups · ~20 minutes · you run the website

Two years from now you manage the e-commerce site of a mid-size retailer (pick one: a regional grocer, a furniture brand, a university bookstore, a B2B parts distributor). Shopping agents are a growing share of your traffic and they do not identify themselves. Choose a posture and defend it:

  • Block: bot management, terms of service, legal letters. You keep the relationship; you also turn away every shopper who delegated.
  • Allow and identify: accept signed agents (Web Bot Auth, Trusted Agent Protocol) with their own rate limits, their own terms and their own checkout path.
  • Join a platform: list through ACP, UCP or a Muse-style partnership. The platform brings identified agents and checkable payments, and takes a seat between you and the customer.
  • Build your own agent: Amazon’s route (Rufus, Buy For Me). Expensive, and your agent will only ever shop at you.

Whatever you choose, write the list of things you would require before an agent may place an order on your site, and the three numbers you would watch in the first month.

Produce: the posture, the requirements list, the three metrics, and one sentence on who pays when an agent-placed order goes wrong.

Strong groups notice that the posture is a business question and the requirements list is not: it looks the same for every posture. Identity of the agent (a signature you can verify, not a user-agent string). Identity of the human behind it, in a form you can hold up later (a mandate, or the wallet’s approval record). A payment credential scoped to this order, so a dispute lands on a token and not on a stored card. Rate limits and a kill switch per agent. Logging good enough to reconstruct the Week 2 evidence line for any order. Clear terms on returns and fraud liability for agent-placed orders, agreed before the first one. The three metrics most groups pick: agent-placed orders as a share of revenue, their return and chargeback rate against human orders, and conversion for agent traffic versus human traffic (which tells you whether blocking would have cost you anything). The liability sentence separates the groups who read page 02: on a stored-card path the merchant and the cardholder argue; on a tokenized path the wallet or network that issued the token holds the record, and usually the first loss.

Open questions, for a thesis or a first job

None of this is settled, which makes it a good place to do original work. Each question below has at least one published starting point.

Markets and strategy
  • Walled gardens or an open web of agents? Rothschild and colleagues (2025) argue that the architecture of agent-to-business communication will decide who captures the margin. Which retailers gain from joining a platform, and which lose the customer relationship?
  • Does ranking fairness survive agents? In Microsoft Research’s Magentic Marketplace simulation (Bansal et al., 2025) frontier models approached optimal welfare only under ideal search, degraded sharply with scale, and showed a first-proposal bias that gave response speed a 10 to 30 times advantage over quality. What does that do to sponsored listings, to late sellers, to price competition?
  • Do agents pay more or less than people? Price discrimination against (or by) agents, negotiation between buyer and seller agents, and the sandboxed “virtual agent economies” Tomašev and colleagues (2025) propose for studying it safely.
Security, trust and measurement
  • How do the protocols fail? A 2026 systematization of agentic-commerce security and a protocol-level attack study of ACP- and UCP-style flows (Louck, 2026) catalogue manipulation of carts, tokens and merchant responses. Which controls on this page stop which attack?
  • Approval fatigue: Muse shows an approval card for every purchase. After how many cards does a person stop reading? What should the card show so that a wrong quantity is caught in under fifteen seconds (Exercise A on page 04)?
  • How good are shopping agents, really? WebShop (Yao et al., 2022), τ-bench (Yao et al., 2024), DeepShop (2025) and UserArena (2026) score task success; Bazaar (2026) scores pricing. A benchmark that scores money lost on wrong purchases, not tasks failed, is a gap worth filling.

Papers: The Agentic Economy ↗ · Magentic Marketplace ↗ · Virtual Agent Economies ↗ · SoK: Security of Autonomous LLM Agents in Agentic Commerce ↗ · Protocol-Level Attacks on Agentic Commerce Platforms ↗ · WebShop ↗ · τ-bench ↗ · DeepShop ↗ · UserArena ↗ · Bazaar ↗

Watch

Stripe Developers: becoming agent-ready with ACP (5 min)November 2025. What a merchant implements to sell inside an agent, in Stripe’s own words.
OpenAI: Buy it in ChatGPT (30 seconds)September 2025, the launch clip. Count the taps between “find” and “paid”, and find the approval.

Longer, optional: Stripe Developers, The future of agentic payments (15 min, March 2026) ↗ · Shopify Engineering, UCP and agentic commerce (40 min, May 2026) ↗.

Trust architecture

Five controls, none of them a prompt

“We told the agent not to overspend” is not a control. Each of these lives in code or configuration, outside anything the model can talk its way past: the Week 10 governance lesson, applied to money.

ControlWhat it stopsWhere it lives
Spending capThe runaway: an agent that misreads a price, a quantity, or a loop and spends far past what anyone intended. Caps come in layers: per transaction, per day, per month.The harness and the payment credential. Enforced before the tool runs, and again by whatever issues the funds.
Merchant allow-listPurchases from places nobody vetted, including the storefront that a poisoned search result invented. If it is not on the list, the call fails.Tool configuration. The agent cannot add to its own list.
Per-transaction mandateScope creep between approval and execution: an approval for one item at one price being reused for a different, larger purchase.The authorization record that travels with the payment. The idea AP2 is trying to standardize.
ReceiptSilent action. Every completed purchase produces an artifact a human can read, tied to the approval that permitted it.Your system of record, plus the customer's inbox.
Audit trailUnanswerable questions later: what did the agent see, what else could it have chosen, why this vendor, who approved it.Trace storage: the Week 2 evidence standard, retained for as long as a dispute could arrive.

The spend ladder

Do not choose between “fully autonomous” and “ask a human every time.” Sort actions into three rungs, and be explicit about the boundary between them.

Auto-runreversible, routine, small

Reading a catalog, checking inventory, pricing options, drafting the cart. Nothing here costs anything to undo, so gating it only teaches people to stop reading approvals. Log it; do not interrupt for it.

Cappedspends, but bounded

A repeat purchase of a previously approved item, from an allow-listed vendor, under a fixed per-transaction and monthly limit. The agent may act, but the limit is enforced in code, and hitting the cap raises an alert rather than silently failing.

Approval requiredirreversible, novel, or large

New vendor, new item, amount above the cap, anything that touches a contract or a subscription. The agent proposes; a named human approves. This rung is small on purpose. A long approval queue is an approval queue nobody reads.

The approval screen is part of the control. A gate that shows “Agent wants to place an order — Approve?” is theatre. Show the decision: the exact items and quantities, the total, the vendor and whether it is allow-listed, the budget before and after, which instruction this purchase is meant to satisfy, and what the agent will do instead if you decline.
Discussion
Discussion questionA customer disputes a purchase their agent made: “I never approved that.” You run the merchant. What do you need to have collected, at the moment of the sale, for that sentence to be answerable rather than arguable?
Three things, and they must have been captured then, not reconstructed later. (1) An authorization record tied to this specific purchase: what was permitted, for how much, for which items, valid until when, and signed by something you can verify rather than asserted by the caller. That is the problem the mandate idea in AP2 is aimed at. (2) A binding between that authorization and this transaction: a payment credential or token scoped to this cart, so an approval for one order cannot be replayed for another; the shared-token design in ACP is aimed at this side. (3) The trace: what the agent saw, what it selected, and the confirmation you returned: the Week 2 evidence standard, kept as long as a chargeback window plus your retention policy. Notice what is not on the list: the agent's own claim that the user approved it. Anything the agent asserts about its own permissions is exactly the thing under dispute.
Concept check

Three questions before you move on

← BackWeek 11 home