TechCrunch reported this week that Mecka AI is closing in on a $500 million valuation in a Sequoia-led round, and the thing they’re selling is people. People wearing rigs, opening cabinets, folding laundry, loading a dishwasher badly and then correctly, all of it recorded so a robot can eventually learn what a dishwasher is.

I don’t work with dishwashers. But the number stuck with me, because it puts a price tag on something the rest of us get for free and then throw in the bin.

The data doesn’t exist, so somebody has to manufacture it

Robots have a genuinely hard problem and I don’t want to be glib about it. The internet is full of text and pictures and video, and almost none of it encodes what it feels like to apply four newtons of force to a cabinet handle that sticks. There’s no scraped corpus of proprioception. There’s no Common Crawl for wrists.

So you build the dataset from zero. You pay humans, you buy hardware, you burn a couple hundred million dollars, and at the end you have a pile of demonstrations that didn’t exist before you paid for them. Half a billion is an honest price for that. The scarcity is real.

Now look at the tab you’re in

Session cookie. Logged in. A half-written reply sitting in the compose box, a spreadsheet with forty rows you sorted by hand, a Jira ticket open in the tab two over, the admin panel where you already clicked through the date filter to get the view you want.

Nobody was paid to produce any of that. You produced it by working, over the last twenty minutes, and it’s sitting right there on your screen with a scroll position and everything.

So why does the agent show up as a stranger?

Because almost every AI agent shipped in the last eighteen months starts in a fresh cloud browser on somebody’s server. Clean profile. No cookies. No history. It has never seen your Gmail, doesn’t know which Salesforce instance is yours, and is about to hit a login wall on the first navigation.

And then it asks you to rebuild the context it just discarded. Grant OAuth scopes. Paste in an API key. Copy the email body over. Re-authenticate the thing you were already authenticated into. Mecka is spending $500M because the context genuinely isn’t there; a cloud agent spends your afternoon because it deliberately started somewhere the context couldn’t follow. We wrote about that gap in more detail in cloud browser agents can’t see your tabs, and the more of these launches I read the more it feels like an ideology rather than an engineering constraint.

Reading the page you’re already on

The alternative is dumber than it sounds, which is probably why it took a while to catch on. An extension that lives in the Chrome side panel doesn’t need a training corpus of your work habits, doesn’t need an integration for every SaaS tool you touch, and doesn’t need permission from a vendor’s API team, because it reads the rendered page in the tab you already have focused, using the session you already established when you logged in this morning.

That’s the whole mechanism. Dassi sits in the side panel and looks at whatever you’re looking at. You’re in a Gmail thread, you ask for a reply, it drafts one that references the actual thread instead of a summary you pasted. You’re staring at an internal dashboard whose CSV export lives behind an enterprise tier, and the numbers are on screen, so it reads them off the screen. You’ve got a vendor form with thirty fields and the answers are scattered across three other tabs. Same idea.

There’s no acquisition cost here. The context was a byproduct of you doing your job. Dassi runs on your own key or your existing ChatGPT login, so the model is whatever you want it to be, and the part that actually matters — knowing who you are and what’s on screen — never leaves Chrome. Nyne raised millions to give agents “human context” and I had the same reaction then. The context is not the hard part. Getting the agent into the room where the context lives is the hard part, and it’s a distribution problem, not a data problem.

Where this argument gets weaker

Physical context compounds. A robot that learns cabinets learns every kitchen. Browser context doesn’t generalize like that; my tabs teach an agent nothing about yours. Mecka is buying a durable asset. I’m describing a free one that expires when you close the window.

The cheap thing is right there

Nobody’s raising $500M to collect what your Chrome profile holds, which tells you something, though I’m not certain what. Maybe that it’s too obvious to fund. Maybe that free context is a crap business model even when it’s the better one.

Either way the robots have a good excuse for starting from nothing. Your agent doesn’t.