Gemini agent mode is Google's name for Gemini completing multi-step tasks for a user, such as browsing sites, filling forms and preparing purchases. Google's help center says Gemini will "always ask for your review and confirmation" before sensitive actions including "making purchases" and "completing financial transactions" (Google Gemini Help). In Chrome, Google says auto browse is "designed to pause and explicitly ask for your confirmation" before a purchase (Google blog).
​

This article covers what Gemini's agent features do, how they handle payments, and how that model compares with agents that pay onchain.
​

Gemini agent mode is a set of Gemini features that act on a user's behalf rather than only answering questions. The agent can open web pages, navigate sites, and work through multi-step tasks, with the user able to watch, intervene, or stop it. Google rolls these features out by product surface, including the Gemini app and Chrome.
​

Google's help documentation describes how users assign tasks and manage workflows, noting that for tasks involving web browsing "Gemini will ask for confirmation first" (Google Gemini Help). Availability depends on plan and region, so the help center is the reference for who can use it.
​

Gemini agent mode can prepare a purchase, but Google says it asks the user to review and confirm before completing purchases or financial transactions. The user stays the final approver. That is a human-in-the-loop design: the agent does the steps and the person signs off on the payment.
​

Human approval is the simplest spend control and the slowest. It works well for occasional purchases and less well for agents that need to pay many times a day, such as paying for API calls. For those cases, limits are usually enforced by the payment rail or wallet instead. See AI agent spend controls and agentic payments.
​

Gemini agent mode and onchain agent payments differ in where the payment happens and who approves it. Gemini works through websites and the user's existing payment methods, with confirmation before purchase. Onchain agents pay from a wallet with stablecoins, and limits are set by wallet permissions or transaction logic rather than a confirmation screen.
​

Gemini and ChatGPT both let agents help with purchases but take different routes. Google emphasizes user confirmation inside browsing tasks. OpenAI introduced Instant Checkout in ChatGPT through the Agentic Commerce Protocol, an open standard co-developed with Stripe for agent-led purchases (OpenAI).
​

See ChatGPT Apps SDK and payments for the OpenAI side and what is agentic commerce for the broader category.
​

Yes. Gemini has agentic features that complete multi-step tasks for the user, similar in spirit to ChatGPT's agent features. Google's help center describes how users assign tasks and how Gemini asks for confirmation before web-browsing tasks and before sensitive actions such as purchases (Google Gemini Help). In Chrome, Google describes auto browse, which navigates sites for the user and pauses for confirmation before a purchase (Google blog).
​

The names and availability differ by product and plan, and Google rolls features out over time. The help center is the reference for current availability in a given country and account type.
​

Gemini's confirmation model keeps the user as the final approver for consequential actions. Google says Gemini will "always ask for your review and confirmation" before actions including "making purchases" and "completing financial transactions" (Google Gemini Help). In practice, that means the agent does the legwork, such as finding the product, filling in details and getting to the checkout page, and then stops and asks.
​

That design has clear strengths and limits:
​

For consumer shopping, the trade-off is sensible. For machine payments, such as paying per API call, a person approving each one does not scale, which is why those cases use other controls.
​

Gemini's agent features suit tasks that involve several steps across websites. Typical examples:

In each case, the user can watch, step in or stop the task. That visibility is part of the safety design, and it is worth using: watching the first few tasks shows where the agent gets confused.
​

Gemini's agent features work through merchant websites and the user's existing payment methods, which are usually cards. Card payments have an important safety property: chargebacks and disputes give the payer a way to recover from a mistake or fraud. Stablecoin payments onchain settle finally, with no chargeback, so limits have to work before the payment rather than after.
​

Neither replaces the other. A consumer assistant buying a jacket is a card problem. An agent paying for data per request is a stablecoin problem. See agentic payments.
​

Human confirmation works as long as the person is present and paying attention. It breaks down in three situations: when the agent must act while the person is away, when payments are too frequent to review one by one, and when the person has started approving without reading. All three push toward limits that the system enforces without asking.
​

For onchain agents, those limits can live in the wallet (only so much is available), in a signing policy (only approved recipients and amounts), or in the transaction (conditions checked against live state). Eco's Sauce SDK covers the last: a program checks conditions such as recipient, amount and minimum received, and reverts if they fail (Sauce docs). The launch post describes the outcome as, "If they hold, the transaction completes. If they don't, nothing moves" (Eco Sauce launch post). That holds per transaction, on one chain. See AI agent spend controls.
​

Gemini and ChatGPT are part of a wider move toward agentic commerce, where AI assistants help people buy. The approaches differ. Google emphasizes confirmation inside browsing tasks across existing sites. OpenAI introduced Instant Checkout in ChatGPT through the Agentic Commerce Protocol, an open standard co-developed with Stripe that merchants integrate so purchases can complete inside ChatGPT (OpenAI: Buy it in ChatGPT). For merchants, that means two paths to agent customers: make the site easy for browsing agents to use, and integrate commerce protocols where they matter. See what is agentic commerce and ChatGPT Apps SDK and payments.
​

When an agent browses on your behalf, it sees what you would see: account pages, order history, saved addresses and sometimes payment details. That makes privacy a practical question, not an abstract one. Before handing an agent a task, consider which sites it will visit while signed in as you, what personal details it will need to fill in forms, and whether you are comfortable with that information being processed to complete the task. Google's help center and privacy settings describe how Gemini handles activity data; review them for your account.
​

A simple habit helps: give the agent only the details the task needs. If it is booking a table, it needs a name, time and party size, not access to your whole inbox.
​

Consumer shopping is the most visible kind of agent payment, but not the largest category of need. Developers are building agents that pay for API calls, compute, data and services from other agents, often in amounts too small and too frequent for a person to approve each one. Those agents rely on wallets, stablecoins and payment standards such as x402, with limits enforced by the wallet or transaction. See x402 explained and agent wallets.
​

Merchants will see more traffic from agents acting for customers. Practical steps include keeping product pages clear and structured so agents read them correctly, making checkout steps predictable, showing totals and terms plainly at the confirmation step, and monitoring for unusual patterns that may be automated abuse rather than genuine customer agents.
​

Google says Gemini asks for review and confirmation before purchases and financial transactions.
​

Google's documentation does not describe crypto payments for agent mode.
​

It depends on plan and region; check Google's Gemini help center.
​

Yes. Gemini has agentic features for multi-step tasks, including browsing in Chrome, with confirmation before purchases and financial transactions.
​

Google describes users being able to watch, intervene in or stop tasks. Check the help center for the controls in your version.
​

Google's design keeps the user as the final approver for purchases and financial transactions, so a person reviews every payment.
​

Google's help center describes confirmation before purchases and financial transactions, which means a person approves each payment. Recurring machine payments, such as paying per API call, are better handled by payment systems built for that, with limits enforced without a confirmation screen.
​

The merchant, the exact item and quantity, the total including taxes and shipping, the delivery address, and any subscription or recurring terms. If anything looks wrong, cancel and correct the request rather than fixing it after payment.
​

Sources: Google and OpenAI pages linked inline, accessed November 2026.
​