Hark previewed a browser-use agent this week, and the demo does the things these demos always do. It searches. It compares a couple of options. It fills a form and books something. Smooth, competent, and every single task happens on a page that any stranger with a URL could load.

That’s not a knock on the engineering. Driving a browser reliably is genuinely hard, and the clip looks better than most. But I keep noticing that the task list in every browser-agent preview reads like a tourist itinerary, and my actual Tuesday does not.

Sort your own tasks into two piles

Take whatever you did in a browser yesterday and split it.

Pile one: looking up flight prices, comparing three vendors’ pricing pages, reading a public changelog, filling out a conference registration form, checking whether a restaurant has a table Thursday. All of this works with zero identity. A fresh cloud browser opens the page, sees the same thing you’d see, and gets on with it.

Pile two: clearing the 40 emails that came in overnight, moving four deals to “negotiation” in HubSpot, pulling last month’s failed charges out of the Stripe dashboard, updating a Jira ticket, exporting a CSV from your analytics admin, cancelling that subscription you forgot about. Every one of these is a page that shows a login screen to anyone who isn’t you.

Pile two is where the hours go. I’d guess it’s 80% of my browser time and close to 100% of the browser time I resent.

Identity, not difficulty

The interesting thing about that split is that it has almost nothing to do with task complexity. Booking a flight involves multi-step navigation, date pickers, seat maps, and a payment form, and it’s genuinely harder as a clicking problem than archiving fifteen emails. But the flight booking runs on the public web and the archiving does not, so a previewed cloud agent can attempt one and simply cannot reach the other.

The boundary is a cookie. That’s it. Your session cookie is the difference between a page that renders your CRM pipeline and a page that renders a login form, and no amount of model capability crosses that line, because the line isn’t made of reasoning.

So you hand over the password?

This is where cloud browser agents get uncomfortable. To reach pile two, a remote browser needs your credentials, which means you’re typing your Gmail password into someone else’s infrastructure and then dealing with 2FA on every fresh session. Some products paper over this with credential vaults. Some ask you to log in inside a streamed remote browser window, which technically works until the session dies and you do it again. I wrote about the mechanics of that in more detail in Cloud Browser Agents Can’t See Your Tabs, and the summary is that the whole thing is a workaround for a problem that only exists because the browser is in the wrong place.

An extension in your own Chrome starts the task already authenticated, because it is your Chrome. Dassi runs in the side panel of the browser you’re already using, so the Gmail tab it operates is the Gmail tab you were reading thirty seconds ago. No password handoff, no vault, nothing to re-auth after a restart. The agent inherits your logged-in state the same way a new tab does. That’s not a clever feature, it’s just what happens when you don’t move the browser to a datacenter.

Preview

Also: it’s a waitlist. Hark’s agent isn’t something you can use on the Tuesday you actually have. OpenAI killed Atlas after everyone decided they didn’t want a whole new browser, and the graveyard of previewed browser agents is getting crowded enough that “we showed a demo” and “you can do your work with this” have become genuinely different claims.

What I’d actually check before getting excited

When the next one of these lands, and there’ll be another next week, the question I care about isn’t the benchmark score. It’s whether the demo would still run if the target site required a login. Most of them wouldn’t, and the ones that would are usually asking you to give up your credentials to find out.

If you want the boring version of this argument with more screenshots, Your AI Browser Agent Can’t See That You’re Logged In covers it. Or skip the reading and install Dassi from the Chrome Web Store, point it at your inbox, and see what an agent that already knows who you are can do with the pile you’ve been avoiding.

I do hope Hark ships something good. I just want it to happen in the tab where I’m already signed in, and not in a pristine cloud Chrome that has to introduce itself to every website like it just moved to town.