AI Browser Agents After Atlas: Claude in Chrome, Gemini Auto Browse, Copilot Mode and Comet

A year ago the pitch was a new browser. OpenAI's Atlas, Perplexity's Comet, a handful of others: replace Chrome with something that has an agent in it, and the agent will do the web for you. Atlas lasted ten months. OpenAI announced in July that it was being wound down, it stopped working on 9 August 2026, and the agent moved back into ChatGPT. The idea did not die; it just stopped being a browser and became a feature of the one you already had.

So the field in September 2026 is one browser and three things that live inside other browsers. We ran the same tasks on each, then read what every vendor says on its own page about what not to do, because on this subject the safety pages are the most useful documentation there is.

What they are for

The tasks these tools are good at are the same on all four, and they are less glamorous than the launch videos. Comparing across tabs: open eight product pages and ask for a table of price, delivery time and return policy. Filling forms: the same address and details into several sites. Collecting: one field from every page in a list. Summarising the page you are on and answering questions about it. None of that is hard. All of it is tedious, and the agent does not get bored on page seven.

What they are not for, by the vendors' own account, is anything involving money or credentials. Google's help page for auto browse says it plainly: you are responsible for Gemini's actions during a task, including mistakes and unexpected results like purchases. Anthropic's page for Claude in Chrome says to avoid financial transactions, password management, or anything involving sensitive personal data. When the people selling the tool put that in writing, believe them.

Claude in Chrome: the most controllable

Claude in Chrome is an extension, available on all paid Claude plans, and its design is built around how much rope you want to give it. It reads the page you are signed in to, then clicks, types and fills forms, and you choose between letting it complete tasks with built-in safeguards or a permissions mode that grants access one site at a time. It stops before sensitive actions such as purchases and asks you. On Team and Enterprise plans, admins can switch the extension on or off for the whole organisation and set allowlists and blocklists of sites, which is the feature that makes it deployable at work rather than just usable at home.

Anthropic says its prompt-injection defences have been tested against real attacks and still tells you to start with trusted sites and familiar workflows. Both statements are true, and the second is the one to act on. If you are already handing whole tickets to a coding agent — we wrote about which ones to hand over — the same instinct applies here: give it the browsing jobs that are tedious and checkable.

The vendors' safety pages agree on the boundary. Inside it — comparing, filling, collecting — the agents are useful. Outside it — paying, logging in, anything you would not want undone — they are not yet, and the vendors are the ones saying so.

Gemini auto browse: the most ambitious, the most restricted

Google's version lives inside Chrome as part of Gemini, and auto browse is the mode where it completes multi-step tasks — the examples on Google's page are shopping, travel booking and reservations, and with Gemini 3 it can look at a picture, find similar items and add them to a basket within a budget. It pauses for what Google calls sensitive steps, including finalising financial transactions and accepting terms of service, asks before sending communications or modifying your data, and will not enter payment details itself. There is a Take over task button on the active tab and a stop icon in Gemini, and the feature is labelled experimental.

The restrictions are the story. You need Google AI Pro or Ultra, a personal Google account rather than a work or school one, to be 18 or over and in the US, and it does not work in Incognito or on iPhone and iPad. It is on desktop and rolling out to select Android devices. If you meet all of that, it is the most capable of the four at long tasks. If you do not, it is not an option this year.

Copilot Mode in Edge: free, and asks first

Microsoft's approach is a mode in Edge rather than a separate product. Turn it on and Copilot can read across your open tabs to compare options and pull out key details, take actions such as searching and filling forms with your approval before anything is finalised, and group your browsing into Journeys you can return to. It is free with Edge, every feature is optional and can be switched off, and for work accounts Microsoft says your data stays in the tenant.

It is the least agentic of the four in practice — more assistant than agent — which is either a limitation or the correct amount of caution depending on how you feel about a browser acting on its own. For someone already on Edge who wants the tab-comparison trick without a subscription, it is the obvious place to start. If you are choosing a browser at all, our browser comparison covers the wider decision.

Comet: still a browser

Perplexity's Comet is the one survivor of the "new browser" wave. It is a full browser on Windows, macOS, Android and iOS with the assistant in a sidebar that knows what is on the page, can run tasks across tabs, and has Perplexity's search engine built in. It launched as a premium product and has spent 2026 becoming free with usage limits, with paid Perplexity plans adding heavier background tasks.

It is also the tool with the most public prompt-injection case study. In 2025, Brave's security researchers showed that instructions hidden in a Reddit post could steer Comet's assistant into acting against the user, and their write-up is the clearest explanation of the problem we have read. Perplexity addressed the specific issue, and the write-up remains the reason we would not let any browser agent — Comet or otherwise — near an account we cared about.

Side by side

	Form	Cost	Availability	Pauses before
Claude in Chrome	Chrome extension	Any paid Claude plan	Paid plans; admin controls on Team and Enterprise	Purchases and other sensitive actions; per-site permissions
Gemini auto browse	Mode in Chrome and Android	Google AI Pro or Ultra	US, 18+, personal accounts, not Incognito or iOS	Financial transactions, terms of service, sending or modifying data
Copilot Mode	Mode in Edge	Free	Edge users; optional	Any action, with approval
Comet	Standalone browser	Free with limits; paid plans add more	Windows, macOS, Android, iOS	Sensitive steps, per Perplexity

How we use them

Claude in Chrome in permissions mode, allowed on a short list of sites we read for research, for the tab-comparison job and nothing else. Copilot Mode on the Windows machine for the same. Nothing is logged in to a bank, a password manager or an email account while an agent is active, which is a rule we keep by running the agent in a separate browser profile with no saved logins — a habit we borrowed from how we set up password managers.

The agents will get better and the restrictions will loosen. What will not change is that a program acting on your behalf on a web page it did not write is reading instructions from a stranger. Until the vendors take the money warning off their own pages, keep it out of the accounts that matter.