I opened Hacker News this morning and counted three browser agent launches on the front page. One was titled “Show HN: I created an open source autoscaling AI browser agent.” Another was an AI Browser Agent Leaderboard scoring these cloud-hosted agents. The third was Yamak, also open-source, also cloud-first.

But the autoscaling framing is what caught my eye. And it’s the quiet assumption underneath all of these: that the bottleneck for AI browser agents is throughput, and the solution is spinning up more headless Chromium instances in some provider’s fleet.

And I only have one browser.

The wrong axis

You don’t need to scale a browser agent. You need a browser agent that can see what you see. The thing I actually want automated is the tab I’m already looking at, logged into Gmail with 2FA already cleared, three filters applied, a half-written draft sitting in the compose window. My personal browser contains a year of saved passwords, a session cookie from a SOC 2 audit vendor that took forty minutes of SSO wrangling to obtain, a LinkedIn tab open since February, and three Notion workspaces I signed into from different laptops. None of that is portable. And none of it scales horizontally.

An autoscaling cloud agent sees none of this. It boots up a fresh Chromium, navigates to gmail.com, and hits the login wall. To do anything useful it either needs my credentials shipped to its server, or it needs me to babysit an OAuth flow every time. Either way, the gain from “scaling to 1000 concurrent browsers” is zero, because the bottleneck was never compute. The bottleneck was identity. And identity doesn’t live in a Chromium container, it lives in the Chrome profile on whatever laptop you happen to be sitting at, with its specific extensions installed and session cookies collected over months of use.

The leaderboard measures the wrong thing

But the leaderboard post got me. It ranks agents by accuracy on WebVoyager and Mind2Web, task sets that run on public websites using test accounts. Which is fine, if you believe the real-world use case for a browser agent is logging into dummy accounts on public websites.

But the tasks that actually matter to me don’t live on public websites. They live on my company’s internal dashboard, my bank portal, the order tracker for the vendor that keeps forgetting to email me updates, and the admin page for a WordPress install I inherited four years ago. A cloud agent with 96% accuracy on WebVoyager can’t see any of those. I wrote about this in the post on what leaderboards miss, but the autoscaling framing sharpens the critique. So scaling a blind worker just gives you more blind workers.

What’s actually hard

A weird thing about this space. The hard technical work in a cloud agent platform is mostly infrastructure: fleet management, browser pool warmup, proxy rotation, CAPTCHA solving, headless detection evasion. That’s real engineering. And I don’t want to be dismissive of it.

But it’s engineering that exists to solve a problem the architecture itself created. You only need proxy rotation because you’re hitting sites from datacenter IPs that look like bots. You only need CAPTCHA solving because the session has no history. You only need browser pool warmup because each run starts cold. But a local browser extension sitting inside Chrome has none of these issues, because it was already there when you were reading your email, and the site has no idea anything unusual is happening.

Limits

So the local approach isn’t magic. It can’t run while my laptop sleeps. It can’t parallelize across fifty tasks. For scraping product catalogs at scale, the cloud model wins. For knowledge work, it loses.

So what does Dassi do

So this is why Dassi is a Chrome extension. It runs in the tab you already have open, with the session you already have, using either your ChatGPT login or an API key you control. And there is nothing for Dassi to autoscale, because your personal browser task throughput is bounded by, well, you. One tab, one human, a few things that need doing. You can install it from the Chrome Web Store.

The autoscaling pitch is aimed at a buyer who thinks they have a bot farm problem. But most of us don’t. We have an inbox problem, an expense-report problem, a CRM problem, and every one of those problems lives behind a login that our actual computer already holds.

Scaling a cloud agent to a thousand headless Chromes will not get you to inbox zero. But a browser agent in your own browser might.