Mark Zuckerberg reportedly told Meta staff that AI agents haven’t progressed as quickly as he’d hoped. That’s a strange thing to hear from the guy who spent a year telling everyone agents would reshape work by now, and it landed in my feed next to three separate startup launches all promising “autonomous agents that just do things for you.” The contrast was almost funny.

Because the models are fine. GPT-5.2 is fine. Opus 4 is fine. If you sit one of these things down with the right information, it reasons through multi-step problems better than most interns I’ve worked with. So when the most agent-bullish CEO in tech quietly admits the timeline slipped, the interesting question isn’t “are the models too dumb.” They’re not. The question is what they’re missing.

What actually stalled

I think Zuckerberg’s admission is really an admission about context, though I doubt he’d phrase it that way.

Here’s the shape of the problem. Almost every agent product you can buy today runs on a rented cloud server somewhere. You type a request, it spins up a fresh headless browser in a datacenter, and that browser knows nothing. It isn’t logged into your Gmail. It isn’t logged into your company’s Salesforce. It hits your bank and gets a login wall, hits your project tracker and gets a login wall, hits basically anything useful and gets a login wall. The agent is brilliant and blind at the same time.

And so the demos that go viral are always the same three tasks: book a flight on a site with no auth, scrape a public webpage, order takeout. Real work doesn’t look like that. Real work happens behind logins, inside sessions you’ve already established, on pages that took you six clicks and a 2FA prompt to reach.

The datacenter has no idea who you are

There’s a version of this problem the industry has been trying to paper over with plumbing. OAuth flows. API tokens. Integration marketplaces where you connect 40 services one exhausting consent screen at a time. The pitch is that if you just authorize the agent to everything, it’ll finally have context.

I don’t buy it, and the security people I trust don’t either. Simon Willison has been banging on about the “lethal trifecta” for a while now, and handing a cloud agent standing OAuth tokens to your entire life is exactly the kind of thing that keeps him up at night. You’re minting long-lived credentials, storing them on someone else’s server, and hoping their breach disclosure is honest. LiteLLM got popped. Proxies leak. The plumbing itself becomes the liability, which is a genuinely dumb way to lose your email.

The deeper issue is that even after all that integration work, the agent still doesn’t have your session. It has a token that approximates a slice of your account. Not the same thing.

Your tab already solved this

Now flip it around. You already have a browser. Right now, as you read this, that browser is logged into everything. Your email is open in one tab. Your CRM in another. Your analytics dashboard, your Notion, your bank, the internal admin tool that has no API and never will. All authenticated. All sitting there.

That’s the context every cloud agent is desperately trying to reconstruct, and it’s already assembled on your machine. This is the whole reason Dassi runs as a Chrome extension in a side panel rather than a server you pay rent on. It uses your real, already-authenticated session. When it opens your inbox, it opens your inbox, because it’s literally your browser doing the opening. There’s no login wall to defeat because you already walked through it this morning.

Nyne raised a pile of money to give agents “human context.” Your browser has had it the whole time. We wrote about that in Your Browser Already Has It, and the point holds: the context problem was solved by whoever built session cookies, decades ago. The agent just needs to run where the cookies live.

Why this is the piece nobody wants to build

So why did the whole industry pour its money into cloud servers, the one place the context isn’t?

Because servers are a business model. You can meter them, autoscale them, charge per agent-hour, put them on a dashboard for enterprise buyers. A Chrome extension that runs on the user’s own machine and uses the user’s own login and, in Dassi’s case, the user’s own LLM key, doesn’t have nearly as many places to attach a price tag. The incentive to build the technically correct thing is weirdly weak.

Zuckerberg’s stall is what happens when a whole field optimizes for the deployment model that’s easy to sell instead of the one that can actually see your screen. The models kept getting smarter. The datacenter kept staying blind. Those two curves don’t meet no matter how many parameters you add.

I’m not convinced Meta will fix this, honestly. Their instinct will be a bigger model and a bigger server, which is more of the same medicine that didn’t work. The fix was never more compute. It was moving the compute to where your logged-in tabs already are.

Anyway. My browser knows who I am. Every agent in a datacenter has to ask.