Google has switched on Gemini in Chrome for every Android user in the US, and brought auto browse — the agent that clicks, types and navigates for you — to phones for the first time.
What arrived, and for whom
The assistant is now open to the full US Android base as of August 18. Auto browse is narrower: on mobile it is limited to Google AI Pro and AI Ultra subscribers, capped at 20 multi-step requests a day for Pro and 200 for Ultra.
That distinction matters more than it sounds. Gemini in Chrome answers questions about the page in front of you. Auto browse actually does things — it opens its own tabs, tagged with an animated AI icon, and works through them using simulated keyboard and mouse input, following links and filling in fields the way a person would.
What it is good at
The sweet spot is the boring multi-step errand. Comparing prices across four retailers. Filling out a long form with details already sitting in your account. Setting up a new service. Pricing a trip across airline, hotel and rail sites. Cancelling the subscription buried three menus deep.
None of that is difficult. It is just tedious and sequential, and a phone screen makes the tedium worse. Handing it to something that never gets bored is a real convenience.
Where it falls over
Long jobs are the obvious weak spot. Anything you would give an intern with "come back in an hour" tends to time out or gets knocked off course by tab activity. This agent wants tasks measured in minutes.
It also struggles on sites with heavy custom rendering or messy markup — roughly speaking, if a page is hard for a screen reader, it is hard for an agent. Sites with strict anti-automation defences will just block it. And for anything you could do in a single tap, calling in an agent is slower than doing it yourself.
The permission problem
This is the part worth slowing down for. Auto browse runs inside your logged-in sessions. Your mail, your bank, a health portal, your employer''s internal tools — all theoretically reachable by whatever the agent decides to do next. Google filters and redacts context when inference happens in the cloud, so credentials and payment details aren''t exposed unnecessarily, and the agent stops at confirmation checkpoints before sensitive actions.
Google''s own documentation is refreshingly blunt about what''s left over: auto browse is experimental, Gemini in Chrome can make mistakes or act unexpectedly — including completing a purchase without permission — and the safeguards don''t guarantee protection against every risk.
That''s a company telling you in writing that its agent might buy something you never asked for.
How to use it without regretting it
Think of auto browse as a capable assistant with shaky judgement. Give it errands, not authority. Actually read the confirmation prompts instead of tapping through them. Keep it away from money and medical records until you''ve watched how it handles low-stakes work. And when a task ends with funds leaving your account, do that step yourself — the minute you save isn''t worth the cleanup.
Image: Bastian Riccardi, via Pexels
