Mark Zuckerberg reportedly told Meta staff that AI agents haven’t progressed as fast as he’d hoped. I saw the line quoted from an internal meeting, and my first reaction wasn’t surprise. It was recognition. Because almost everyone I know who has tried to hand real work to a general-purpose agent has hit the same wall, and it is not a wall built out of model intelligence.

The models are fine. GPT-5.2 is fine. Gemini 3.1 Pro is fine. The problem is where they are standing when you ask them to do something.

An agent in the cloud is an agent locked out of your life

Picture what a typical cloud agent actually has when it spins up. A fresh browser. No cookies. No sessions. No memory that you’re the person who pays for three of the SaaS tools it’s about to try opening. It lands on a login screen and stops, or it tries to log in and trips a bot check, and now you’re solving a CAPTCHA for your own assistant.

That’s the gap Zuckerberg is describing without naming it. Not “the reasoning isn’t good enough.” The reasoning is great. The agent just can’t get into the room where your work happens.

What he’s actually admitting

So my read is this. Meta, OpenAI, Google, all of them poured enormous effort into agents that live on servers and reach out to the world through APIs and headless browsers. That architecture demos beautifully. It falls apart the second the task requires being you: logged into your Gmail, your Notion, your company’s internal dashboard that has no API and never will.

And most real work is exactly that. It’s not “book me a flight” in the abstract. It’s “reply to the four emails in my inbox that are actually urgent,” which requires seeing my actual inbox, with my actual threads, while I am actually logged in.

The browser already solved the context problem

I wrote about this a while back, when a startup called Nyne raised money to build a “human context layer” for agents. The whole premise assumed context was a thing you had to construct and pipe in from the outside. But your browser is already the context layer. It holds your logins, your open tabs, your half-filled forms, and the dashboard you were squinting at ten minutes ago.

A browser-native agent like Dassi doesn’t spin up a clean cloud sandbox and start from zero. It works inside the Chrome tab you already have open, using the session you already authenticated. When it drafts a reply, it drafts it in your real Gmail. When it pulls numbers off a vendor portal, it’s your logged-in portal, not a stranger’s locked door. Nothing gets re-authenticated, nothing gets shipped to a remote server to be poorly reconstructed. This is a boring architectural detail, and it is also the entire difference between an agent that disappoints and one that does the thing you asked.

Why the big labs keep missing this

Because the incentive is pointed the wrong way. An agent that runs on your servers is an agent you can meter, monitor, and monetize. An agent that runs inside the user’s own browser tab, using the user’s own login, on the user’s own machine, is architecturally the opposite of a data funnel, and it does not fit the business model that funds these labs in the first place.

So they build the cloud version. It benchmarks beautifully. And then it disappoints in the one place it was supposed to help, and a very smart CEO stands up in an all-hands and says the progress isn’t what he hoped.

Damn right it isn’t. He’s measuring a car by how fast the engine spins while the wheels are off the ground. I’ve argued before that cloud browser agents can’t see your tabs, and that limitation is the whole story here, dressed up as a mystery.

The version that’s already in the room

If you want to feel the difference, stop reading about agents and watch one work in a tab you are already signed into. Dassi runs as a Chrome extension in your side panel, reuses your existing sessions, and lets you bring your own model key or just log in with the ChatGPT subscription you’re probably already paying for. It’s on the Chrome Web Store.

Zuckerberg’s disappointment is real. It’s just aimed at the wrong layer. The agents were never dumb. They’ve been standing outside your house describing the furniture through the window, and someone finally handed one a key.