OpenAI Wants to Sell You a Phone. The Browser Already Won.
The phone has the wrong-layer problem. The leaked OpenAI/Jony Ive device keeps getting framed as “no apps, just agents,” and every time it surfaces the framing assumes apps are still the dominant computing surface that needs replacing. But most of the apps people use all day are already browser tabs wearing different clothes.
Slack is Chromium. Notion is Chromium. Linear is a browser. Figma runs in a browser. Even Spotify on the desktop is basically a webview with the chrome stripped off. So the question isn’t whether agents should replace apps. That fight is mostly over. And the answer to “where does the agent sit” has been visible for a while: it sits in the tab, because the tab is where the work already lives.
The browser already won the layer fight
I keep thinking about the old argument that the browser would eat the OS. Mostly correct. ChromeOS is the proof of concept, but more importantly the SaaS wave made every knowledge worker into someone who lives in 30 open tabs and a Calendar invite. So the phone is not where work happens. It’s the place where you check on work that happened somewhere else.
And building a phone where AI replaces apps is solving a 2014 problem. Apps have been losing to web wrappers for a decade. So the interesting question now is what happens to the browser when the agent shows up, not what happens to the phone.
What an agent inside a tab actually does
Take Gmail. The version of “AI for email” the OpenAI phone implies is an assistant that reads your mail through an API, drafts responses, and pings you on a device. But that assistant has to be granted OAuth access, has to handle attachments through a server somewhere, has to deal with rate limits, and has to convince you it isn’t training on your inbox. So the version that actually works today is dassi sitting inside your Gmail tab, reading what you can read, drafting in the compose box you already have open, sending from the account you’re already logged into. No OAuth dance. No copy-pasted threads. No “permission to access your contacts.”
And this is the boring part most agent narratives skip. The web is full of authenticated state, like your Linear board, your bank, your Stripe dashboard, your customer’s Salesforce, and that state is already loaded in the tab you have open right now. An agent that lives in the browser inherits all of it for free. But an agent that lives in a phone has to rebuild every connection from scratch through partner APIs, with permission scopes the user has to approve one at a time, and with a security model whose surface area is honestly kind of frightening. So that’s a vastly bigger attack surface for a vastly smaller payoff. We covered the same logic in Cloud Browser Agents Can’t See Your Tabs, and it applies to a phone agent too, except worse, because the phone doesn’t even have a tab to begin with.
Apps were never the canvas
The canvas is the page, and most knowledge work now happens across 30 of them, plus a calendar, plus whatever password manager you can’t bring yourself to switch away from.
OpenAI’s real product
Hardware lock-in. If agents replace apps and the agents live inside your phone, then the phone vendor controls the agent layer, which controls every transaction you make online. So that’s a wildly valuable position to claim. And it’s why Sam Altman keeps showing up at Foxconn, and why Apple is reportedly panicking internally about being the second mover. But none of that changes the fact that for actual users right now the agent with the most context and the fewest permission walls is the one running where they already work, which is, depressingly often, a Chrome tab open to a SaaS dashboard at 11pm.
So you can install dassi from the Chrome Web Store and it already does most of what the OpenAI phone is supposed to do. The phone doesn’t exist yet. The tab does.
Maybe the phone ships and I’m wrong
It’s possible the device is so good it pulls behavior off the laptop and into a pocket form factor. I doubt it. Because the work people actually want agents for — invoicing, scheduling, sales prospecting, research, expense reports, drafting long emails — is not work that fits comfortably on a 6-inch screen, agent or no agent. But Jony Ive has surprised people before.
Either way, the phone is mostly a distribution play. And the agent play has been the tab the entire time. Building a new device to host agents in 2026 is a bit like building a new cassette deck in 2003 because you’ve heard music is going to be a big deal.