> ## Content Index > Fetch the complete content index at: https://www.livain.com/llms.txt > Use this file to discover other available public pages before exploring further. # OpenAI built a whole browser for the agent. The agent didn't need one. - URL: https://www.livain.com/blog/openai-built-a-whole-browser-for-the-agent-the-agent-didnt-need-one/ - Published: 2026-08-27T05:11:20.000Z - Updated: 2026-08-27T05:11:20.000Z - Description: Claude Cowork just got a built-in browser, seventeen days after OpenAI switched off Atlas. The improvement isn't the browser — it's that there are finally two modes, and knowing which job belongs in which one. - Author: Remco Livain - Tags: AI, Marketing Tech, Tooling Anthropic shipped a browser inside Claude Cowork this week. Not an extension, not a separate app — a side panel that opens when a task needs a website, navigates it, reads it, clicks and types, [while you keep working on something else](https://claude.com/blog/cowork-built-in-browser?ref=livain.com). Seventeen days earlier, OpenAI switched off Atlas, the browser it had built for exactly this job. The easy read is that one company got it right and the other got it wrong. That read is wrong, and the real lesson is more useful to anyone deciding what to put in front of their team. ### What Atlas actually got wrong Atlas launched in October 2025 and [was shut down on 9 August 2026](https://techcrunch.com/2026/07/09/openai-is-shutting-down-atlas-but-its-ai-browser-ambitions-are-still-growing/?ref=livain.com), under nine months later. It never shipped beyond macOS. It made no dent in Chrome's share. Internally it got filed under the side quests OpenAI decided to stop running. But OpenAI did not conclude that agentic browsing was a bad idea. It [moved those capabilities into ChatGPT and Codex](https://help.openai.com/en/articles/20001371-evolving-atlas-into-chatgpt-for-browser-based-agentic-work?ref=livain.com) instead — including a built-in browser in its own desktop app. So the difference between the two companies is not that one has an agent-driven browser and the other doesn't. Both now do. What Atlas got wrong was the ask. It required me to move in. To leave Chrome, where my sessions, my extensions and fifteen years of muscle memory live, and adopt a new home for all of my browsing, in exchange for AI help with a slice of it. That's an enormous switching cost for a narrow benefit, and I never made the trade. Almost nobody did. > As Brad Lightcap of OpenAI put it on 20VC: there are only 1–3 things that are relevant for a company — not ten. ### The improvement is that there are now two modes, not one ![Line drawing of two identical browser windows: the left one used by a person at a keyboard, the right one standing alone and working by itself.](https://storage.ghost.io/c/2d/96/2d9620e3-6a0c-455c-9217-68a5989a3c45/content/images/2026/08/modes.jpg) Two jobs, two browsers: the one I'm signed into, and the one that works without me. Here's the part I find good, and it isn't the browser itself. It's that Anthropic named two different jobs and refused to collapse them into one product. Claude in Chrome is for *my* browser — the page I already have open, signed in as me, doing something I'd otherwise do by hand. Updating a record, working through an inbox, fixing the document in front of me. The built-in browser is for *a* browser. Gathering research, pulling this month's invoices out of a supplier portal, working through a system that has no connector and never will. Those tasks don't need my tabs. They need any tab, and they need me not to be sitting there watching it happen. That distinction sounds small until you've spent an afternoon with an agent that hijacked the machine you were trying to work on. Handing over the web part of a task while you carry on with the rest is the whole point. Atlas couldn't offer that, because Atlas *was* the machine you were working on. ### The login model is the quiet win The built-in browser is separate from mine. Claude doesn't see my tabs, my bookmarks or my passwords. If I want it signed in somewhere, I carry that login over one site at a time — and banking, email and single sign-on are left out unless I explicitly include them. That is a real answer to a real problem. Half the work I do for clients sits behind a portal with no API, a login that isn't mine to share, and an export button that produces a PDF. I've [worked around missing access before](https://www.livain.com/blog/no-login-to-the-tag-manager/), and the workaround is always slower than having the access. A browser I can grant three specific logins to, and nothing else, is a permission model I can actually explain to a client's IT lead without them going pale. It also fixes a thing that has quietly annoyed me for months: [automations that can only run when I happen to be at the machine](https://www.livain.com/blog/i-scheduled-a-weekly-check-that-could-never-run/). A browser that lives in the desktop app, drivable from my phone as long as that app is open, moves a whole class of recurring web chores from "when I get to it" to "already done." ### What hasn't changed Prompt injection. A page can carry instructions aimed at the agent rather than at me, and the separate browser doesn't fix that — it runs the same classifiers as the extension, and Anthropic's own research still calls the problem [far from solved](https://anthropic.com/research/prompt-injection-defenses?ref=livain.com). Their advice is to [start on sites you trust](https://support.claude.com/en/articles/12902428-use-claude-in-chrome-safely?ref=livain.com), which is exactly right and exactly the sort of thing people skip. So my rule is boring. Known portals, known vendors, known shape of task first. Not the open web, not a site I landed on from a search result, not anything where a wrong click costs money. ### The judgement is which browser, not whether This is the pattern I keep running into. The tool arrives, everyone asks whether to use it, and the useful question is which job it's for. Knowing that a task belongs in my authenticated session versus a disposable one, and that the second kind can run without me, is worth more than any feature on the announcement page. I've watched [ten minutes of the right question kill two weeks of the wrong work](https://www.livain.com/blog/ten-minutes-of-querying-killed-the-content-sprint/) often enough to trust that. OpenAI spent nine months and a whole browser learning that the browser was the wrong container. Anthropic shipped a side panel. Same conclusion, cheaper tuition. ### Sources & further reading **External:** [Claude gets its own browser in Cowork](https://claude.com/blog/cowork-built-in-browser?ref=livain.com) · [Piloting Claude in Chrome](https://claude.com/blog/claude-for-chrome?ref=livain.com) · [OpenAI: evolving Atlas into ChatGPT](https://help.openai.com/en/articles/20001371-evolving-atlas-into-chatgpt-for-browser-based-agentic-work?ref=livain.com) · [TechCrunch on the Atlas shutdown](https://techcrunch.com/2026/07/09/openai-is-shutting-down-atlas-but-its-ai-browser-ambitions-are-still-growing/?ref=livain.com) · [Anthropic on prompt injection defenses](https://anthropic.com/research/prompt-injection-defenses?ref=livain.com) · [Using Claude in Chrome safely](https://support.claude.com/en/articles/12902428-use-claude-in-chrome-safely?ref=livain.com) **Related posts:** [No login to the tag manager](https://www.livain.com/blog/no-login-to-the-tag-manager/) · [I scheduled a weekly check that could never run](https://www.livain.com/blog/i-scheduled-a-weekly-check-that-could-never-run/) · [Ten minutes of querying killed the content sprint](https://www.livain.com/blog/ten-minutes-of-querying-killed-the-content-sprint/)