A post titled “AI automation is quietly de-skilling white-collar workers” sat near the top of Hacker News for most of Monday, and the comments were brutal. Accountants saying they forgot how to reconcile without their AI tool. Junior developers admitting they cannot debug code they did not write themselves. One person described watching a colleague try to draft a client email without GPT and fail so completely it got escalated to their manager.

The anxiety is real and it is not new. Cal Newport talked about this on his podcast a few weeks back, calling it “competence atrophy,” which sounds clinical enough to be scary. But the framing bugs me. People keep treating deskilling like an inevitable side effect of using AI at all, when really it is a side effect of how most AI tools are designed.

The black box problem is architectural

Most AI productivity tools follow the same pattern: you describe what you want, it disappears into a cloud, something comes back. Maybe it is a drafted email, maybe a formatted spreadsheet, maybe a summary of a 40-page PDF. You get the output but you never see the process that produced it, and over weeks and months of this, the process fades from your memory because you stopped doing it.

I watched this happen to myself with expense reports. Our company switched to an AI tool that auto-categorizes receipts and fills the whole report. Three months in, finance asked me to manually categorize something the tool rejected and I sat there staring at the form like I had never seen it before. That’s deskilling in miniature. Not dramatic, not career-ending, just this slow erosion of knowledge you used to have because a tool quietly removed the need for it.

Working alongside instead of instead of

Browser agents flip the architecture. Instead of sending your task off to some backend and waiting for a result, a browser agent operates right there in your Chrome tabs, clicking the same buttons you would click, reading the same pages you would read, filling in the same forms.

And the difference, which sounds subtle but changes everything about the experience, is that you watch it happen. You see each step. You can stop it mid-action if something looks wrong, which means your brain stays engaged with the actual workflow instead of atrophying because you outsourced it to a text box in a separate window.

Dassi does this from Chrome’s side panel. It sits next to whatever you are working on and acts within the page, using your logged-in sessions and your visible context. So when it drafts a reply in Gmail, you see it type into the compose field on the same thread you were reading. When it pulls data from a dashboard, it navigates to the dashboard in your tab and you can see exactly which filters it applied and which numbers it grabbed.

The tedium is what should go

Deskilling happens when automation replaces understanding. But not every part of a workflow requires understanding. Nobody becomes a better analyst by manually copying 200 rows from a web table into a spreadsheet. Nobody learns anything from clicking through five dropdown menus to apply the same label to thirty emails. That is mechanical repetition and losing the ability to do it is not a loss.

The tasks you are still doing manually in your browser fall into two buckets: stuff that builds competence and stuff that just burns time. A well-designed agent handles the second bucket while keeping you present for the first.

I keep coming back to a specific example. When I research competitors, the interesting part is reading what they ship and figuring out what it means for our positioning. The crap part is opening twelve tabs, scrolling past cookie banners, finding the pricing page, locating the feature comparison table, copying it somewhere useful. A browser agent that handles the navigation and extraction while I focus on the analysis is not deskilling me. It is removing the parts that were never a skill in the first place.

Staying in the loop is the whole point

The cloud browser agents that spin up a remote Chrome instance somewhere and do everything offscreen are basically the same black box problem wearing a different hat. You do not see what they click, you cannot tell if they misread a page, and the output arrives stripped of all the context about how it was produced.

Working in the user’s own browser solves this almost accidentally. Because the agent has to operate visually, inside your session, on your screen, you end up supervising it the way you would supervise a new hire sharing your screen. You learn from watching it, you correct it when it goes wrong, and you retain the knowledge of how the workflow actually works because you are literally watching the workflow execute in front of you.

Nobody is going to stop worrying about deskilling anytime soon, and honestly some of that worry is healthy. But the solution is not to avoid AI automation entirely, because the people who do are just going to drown in the busywork that their peers automated six months ago. The solution is picking tools that keep you close enough to the work that your skills stay intact while the tedious parts disappear. A browser agent that works in your tabs, where you can see everything it does, is about as close as you can get right now.