Agent Experiences

Atlas Is Dead, Agentic Browsing Isn't

By Alan WeibelPublished 9 min read

The shutdown

OpenAI launched the Atlas browser on October 21, 2025, announced its shutdown on July 9, 2026, and switched it off on August 9, 2026, less than ten months later. What Atlas could do moved into ChatGPT itself: the desktop app, a Chrome extension, and Codex, OpenAI's coding assistant. OpenAI's view was that browsing belongs inside the assistant as a feature, rather than in a separate app people have to switch to.

Source: TechCrunch, July 2026

Atlas got the headlines, but it wasn't the only one to go. In the same ten months, Microsoft retired Copilot Mode in its Edge browser, Google shut down Project Mariner, and Anthropic turned its Chrome extension into something quite different. Taken together, they look less like four companies giving up on agentic browsing (an AI assistant that opens web pages and works through them for you) and more like four companies reaching the same conclusion about where it belongs.

That conclusion has a plain, practical effect on anyone who runs a website, and that's why this matters to you beyond industry gossip. Below is what happened, why, and what it changes about the visitors arriving at your site.

What happened, in order

OpenAI Atlas. OpenAI launched Atlas on October 21, 2025. It was a separate browser, built on Chromium (the open-source code underneath Chrome), with ChatGPT built in and an agent mode that could work through pages for you. OpenAI announced the shutdown on July 9, 2026, and turned it off on August 9. The features survived and moved into the ChatGPT desktop app, a Chrome extension, and Codex. OpenAI's own explanation was that people wanted ChatGPT to be able to use the browser they had, and had little interest in a new one.

Microsoft Edge Copilot Mode. Microsoft retired Copilot Mode on May 13, 2026, and built its AI tools straight into regular Edge, as Business Standard reported. Microsoft had the easiest route of anyone to getting a browser in front of people, since Edge comes with Windows, and it still decided that a separate AI browsing mode was the wrong home for this.

Google Project Mariner. Mariner was Google's research prototype of an agent that operates websites for you. Google's own Mariner page says the service ended on May 4, 2026, and points people to Gemini Agent instead, as Gigazine reported. It's the same pattern again: the experiment ends, and what it could do ships inside products people use every day.

Anthropic's Chrome extension. On August 12, 2026, it became a full client for Claude Cowork, Anthropic's assistant that carries out multi-step tasks for you. So instead of a browser of its own, Anthropic gave its assistant more to do inside the browser you run now.

Perplexity Comet. Comet is the exception, and it shows what it takes to buck the trend. Perplexity made Comet free on macOS, Windows, Android, and iOS, with the iOS version arriving on March 18, 2026. Within 48 hours the iOS app reached number three in the US App Store. It's the one standalone AI browser with a sizeable audience, and it got there by being free, on every platform, with heavy promotion. Comet's shopping agent is also exactly what Amazon took Perplexity to court over, which we cover in the court just ruled your agent is you.

Why the standalone browsers lost

Standalone AI browsers mostly failed as products. Agentic browsing did well as a feature inside assistants people were using anyway.

Atlas worked, so the technology wasn't what held these browsers back. The trouble is that a browser is one of the hardest pieces of software to get anyone to give up. It holds their passwords, extensions, bookmarks, autofill, open sessions, and years of habit. An AI browser had to beat Chrome plus your assistant by enough to justify moving all of that. None of them did, because the assistant could simply be added to Chrome.

What people wanted was smaller: to tell the assistant they talk to every day "look at this page," "book this," "compare these three," or "fill this in." That's a request for the assistant to reach further, and it has little to do with browsers. Every company except Perplexity shipped it as a feature in the end, and the ones that had built a separate browser for it shut that browser down.

What it means for your site

Where the agent lives decides what shows up at your server, and the agents have just moved into the browsers your visitors use every day. Five things follow from that.

Agent traffic arrives inside an ordinary browser. You'll see Chrome running on someone's own device, where you might have expected a script fetching pages from a data center. It comes with the user's cookies, their screen size, their logged-in session, and the same connection fingerprint as any other browser (the technical details a browser shares when it opens a secure connection, which security tools use to tell browsers from scripts).

It carries an ordinary user-agent string. Every request a browser makes includes a short line of text naming the browser or bot that's asking. Agents that run as an extension or inside the browser don't change that line to announce themselves. So you won't see ChatGPT-User, the name ChatGPT gives when it fetches a page because someone asked it to. As far as your server can tell, the request comes from Chrome. Agents that do name themselves still exist and still matter. Nobody has published a count, but they're likely to be a smaller and smaller part of what agents do as more of them move into browsers.

There's nothing in the request to tell it apart. Often there's no field you could build a rule on. Your bot management tools were built to answer the question "is this an agent?", and for this kind of traffic there's no reliable answer at the moment the request arrives.

You can't block it without blocking your customer. Sometimes heuristics (rules of thumb) can spot it from patterns in behavior, timing, or the way someone moves through a page. Even then, the thing you'd be blocking is your logged-in customer's assistant doing what your customer asked. You'd mostly be creating support tickets for yourself.

It won't show up in your bot analytics. Whatever your dashboard says about traffic from AI crawlers (the programs AI companies send out to fetch pages), it misses agents working inside browsers, and every agent that moves into a browser adds to what it misses. You have a measurement problem before you have a policy problem. Since you can't count these visits, behavior is the only signal left: the shape of a session, the path someone takes through your pages, and sequences of steps that look like a task being worked through. Our measurement guide covers what you can still count.

What to do now

  • Stop relying on user-agent strings for your bot policy. They still work for crawlers and fetchers that announce themselves. They can't see agents living inside a browser at all, and the browser is where agents are moving.
  • Set limits and permissions per account or per session, because an account still identifies your customer when the software can't be identified.
  • Assume some of your logged-in sessions have an agent helping. You can't measure how many. Build flows that cope with it: clear confirmations, actions that can be undone, and pages that make it obvious what state things are in.
  • Make your pages make sense to something that can't see them. An agent reading your page does better with proper HTML structure (headings, lists, and buttons marked up as what they are), labels on every form field, content that's in the page when it loads, and forms that do what they say. That work pays off whether the agent is feeling its way through or calling a tool you've published for it.
  • Don't build your plans around any one AI browser. In ten months this category lost three major products, so plan for what agents can do rather than for a particular product.

Why sites now need to tell agents what they can do

If agents work inside the user's browser, in the user's session, looking exactly like the user, then spotting them and deciding what to do about them stops working. You can't set a policy for traffic you can't identify.

What can work is the site speaking up: the site tells the agent what it's allowed to do, and the agent calls a function the site has published rather than pretending to click buttons. That's the case for WebMCP (a proposed browser standard that lets a web page offer its own tools to AI agents) and anything like it. A browser-side API (a set way for software to ask your site to do something) for agents looked like a niche idea in 2025, and it looks like the obvious answer now. We explain what WebMCP is and what it asks of you in WebMCP: the web finally gets an agent API.

A US federal appeals court, the Ninth Circuit, reached the same place. In the Comet case, it reasoned that an agent following a user's instructions is the user acting, so blocking it is closer to blocking your customer than to blocking a scraper. On the technical side, an agent following a user's instructions now looks exactly like the user, so you couldn't block it separately even if you wanted to. The law and the technology got to the same answer from opposite directions within about four months: the browser shutdowns ran from May to August 2026, and the Ninth Circuit ruled on August 4, 2026.

So the useful question has changed. Instead of asking how to keep agents out, ask what you want an agent to be able to do once it's on your site.

References

See the path on your own site

Whichever browsers survive, they will work your pages the same way. See where an agent gets through one of yours and where it stops.

Measure what you can still see

Agent sessions are getting harder to identify. Our measurement guide covers the signals that still work, and the ones that never did.