How far should software be allowed to buy things in place of people. Three announcements offering their own answers to that question came out over three days early this week. Google unveiled an open e-commerce standard, the Universal Commerce Protocol (UCP), at a National Retail Federation (NRF) event, Naver disclosed a shopping agent set for release in the first quarter, and Anthropic opened Cowork, a PC agent for non-developers.
The formats of the announcements differ, but all three touch the same point. When an agent executes on a user's behalf, how is that authority proven and who takes responsibility.
UCP was built together with retail companies such as Shopify, Etsy, Wayfair, Target and Walmart. More than 20 partners including Adyen, American Express, Best Buy, Mastercard, Stripe and Visa have expressed support. Google released the specification as open source. It is a choice to set the reference point by having the standard widely installed rather than monopolizing it.
Technically, UCP does not stand alone. It works in conjunction with AP2, the payment protocol Google announced in September 2025, and it is also compatible with A2A, the specification for agents talking to each other, and MCP, the specification for models accessing external tools and data. At the core is AP2's "Mandate." It is a cryptographically signed contract with tamper-proofing, leaving evidence of what the user instructed and holding the products and prices an agent presents so they do not diverge from the seller's actual information.
UCP's payment authorization works on the same principle. Every authorization is handled through cryptographic proof based on user consent, and payment companies can attach modular processors. A condition worth noting is that a company using this standard still manages its own business logic directly and bears its responsibility as a seller as before. The line that the platform will not swallow the transaction has been written into the standard document.
Deployment plans were also announced. U.S. retailers will be able to list products and services on Google Search's AI Mode and the Gemini app screens, and let customers complete payment right on that screen while searching. Google Pay handles payment and shipping details are pulled from Google Wallet, with PayPal to be added later.
Google did not disclose a start date. Separately, Google has placed brand-specific AI business agents in search results, and Reebok, Lowe's, Michaels and Poshmark are using them.
Several similar attempts are already out. OpenAI has teamed up with Stripe and Shopify to build ACP, which finds products inside ChatGPT and connects through to payment, and Microsoft has released "Copilot Checkout," which pays directly from the Copilot chat window, through Shopify and Etsy routes. The specification names differ, but what they are competing over is the same.
It is whose company's grammar commerce gets recorded in. But no account has emerged of how far these specifications interoperate with each other and where they collide.
The answer Naver put out the same day aims at a slightly different place. It is applying "Agent N," its autonomous AI strategy, to search, shopping, Webtoon and Maps, and will release a shopping agent in the first quarter. Set a budget and hand over the grocery shopping, and it draws up an ingredient list, compares prices, proposes a cart, and continues through after-sales service following payment.
On top of this, for actions that are hard to reverse such as payment, reservations and deleting posts, a user confirmation step is always inserted and final approval is left with the person, and the principle of disclosing the grounds on which the AI reached such a conclusion has been fixed in a document. "Agent N for Business," aimed at small merchants, comes with it.
Anthropic's Cowork chose the side of not hiding the risks. It is a tool that pulls Claude Code, which started as a command-line tool in November 2024 and expanded to the web in October 2025 and to Slack two months later, out in front of people who are not developers. It goes into the desktop app and reads, edits and creates files in designated folders, and through Chrome integration it also handles browser and external app information.
The company itself wrote that because it is designed to work continuously without user intervention, when instructions are vague or conflict with each other, files may be deleted or prompt injection, in which sentences hidden in external documents are read as commands, may occur. It is currently at the research preview stage and available only to subscribers of the top-tier plan.
Placing the three announcements side by side reveals where the contest has moved. Beyond how smart the model is, it has moved to in what form the user's intent is proven, where execution authority is cut off, and who pays when something goes wrong. Google has set its bolt at cryptographic proof, Naver at a human confirmation step, Anthropic at risk disclosure, each at a different point. As for how sellers, platforms and payment companies divide responsibility when a misdelivery or a wrong purchase actually happens, no rules have come out from any side for now.
For the time being, what this technology changes is the time it takes to fill a cart. The dozen-plus minutes spent comparing prices, moving between windows and entering card numbers shrinks to a single sentence. In exchange, users gain one new habit.
It is the habit of checking what the agent chose and why, and stopping it once just before payment. How Naver's shopping agent actually implements that pause when it arrives in the first quarter of this year becomes the first test of whether this approach finds a place in daily life.
