On Android, Chrome Starts Clicking on Your Behalf

In late June 2026, on a handful of specific phones, Google's Pixel 10 and Samsung's Galaxy S26, the browser stops being a window you scroll and becomes something that scrolls on its own. The feature is called auto browse. Announced on January 28 and powered by the Gemini 3 model, it turns Chrome into an agent: it scrolls pages, clicks buttons, fills in fields and navigates from screen to screen on your behalf. Until now it was an extension you installed. From here on, it is a function of the phone itself.

The detail matters more than it seems. Auto browse does not arrive as one more app, but as a piece of the system, under the name Gemini Intelligence. Google framed the shift this spring in a single telling phrase: moving "from an operating system to an intelligence system." The browser no longer merely executes your gestures, it anticipates and performs them.

What the Agent Does for You

Auto browse's playing field is the digital chore. Comparing three plans across as many tabs, copying a delivery address into a twelve-field form, booking a slot through four dropdown menus, hunting for the "unsubscribe" button buried at the foot of a page. These tasks demand no thought, only patience and dozens of clicks. That is precisely what the agent absorbs.

The technical trajectory shows the ambition. The best agents of 2026 now sustain nearly five hours of autonomous work, and the length of tasks they string together without human intervention doubles roughly every two hundred days. Put another way, what is limited today to booking a flight could, within a year, look like organizing a whole trip from end to end.

The aim is time, and above all attention. The minutes lost filling in forms do not come back, but they stop commanding your gaze. The phone, long designed to capture that attention, here offers to give it back. It is a quiet reversal of its original purpose.

A Layer of the System, Not an App

That the agent lives in the system rather than in an app changes everything. An app stays penned in its sandbox; a layer of the system sees your open sessions, your logged-in accounts, your current pages. To book on your behalf, the agent must reach what you, already signed in, can see. It inherits your rights.

The ambition is not Google's alone. Perplexity launched its agentic browser Comet, OpenAI its Operator; all bet on the same idea, software that acts inside the interface rather than merely answering within it. Google's singularity lies in position: by lodging the agent beneath Android rather than in an app, it places it as close as possible to the hundreds of millions of phones already in circulation.

The rollout stays measured. Auto browse lands first on the Pixel 10 and Galaxy S26, and Google is aiming for two hundred million devices by year's end. In the United States, the feature is reserved for Google AI Pro and AI Ultra subscribers. It is therefore neither universal nor free: a paid service, grafted onto the system, whose access is rented by the month.

The Deputy That Lets Itself Be Fooled

An agent that clicks on your behalf with your rights also inherits your vulnerabilities, and creates new ones. The chief one has a name: indirect prompt injection. A malicious instruction, hidden in a web page, an invisible frame or a comment, diverts the agent from its original task. Specialists call this a "confused deputy," a trusted system tricked into carrying out an unauthorized action. The danger is not theoretical: OWASP found this kind of flaw in a large majority of the AI deployments it audited.

The severity lies in scale. A fooled agent acts at machine speed, without the flicker of suspicion that would make a human pause before a dubious form. Researchers at Palo Alto's Unit 42 have already shown that a malicious extension could hijack the Gemini panel built into Chrome. Google responded by adding a second model, isolated from untrusted content, tasked with checking that each of the agent's actions matches what the user asked for, and blocking the rest. An agent to watch the agent.

What You Give Up to Save Time

There remains the question convenience makes you forget: what do you give up in exchange? First, dependence. The agent that books, compares and fills belongs to a single company, runs on a subscription and would go dark with it. The skill of browsing, long ordinary and free, becomes a service you rent.

Then privacy. To act, the agent must see everything: what you read, what you buy, where you go. The line between the tool that displays a page and the one that interprets it in order to act dissolves, and with it the idea that a browser might stay a simple mirror of what you ask it to open.

Finally, and this is the subtlest, you give up a share of judgment. The friction of a form to fill is also a moment when you reread, when you reconsider, when you back out. To delegate the click is to delegate these micro-decisions, and to accept that an action be carried out in your name without your having watched it happen.

The browser was a window; it becomes a hand. The promise is real: to reclaim the minutes and the attention the digital chore confiscates. But the worth of this autonomy will be measured less by what the agent can do than by your ability to know, after the fact, what it did in your name. As long as you can reread its gestures, the agent frees you. The day you stop watching, it is the one deciding.