Someone posted “Botsitting: The Unpaid Labour Behind Every AI Productivity Claim” to Hacker News this morning, and by lunch it was sitting a few slots above the Andrew Bosworth story about how the time AI saves you ought to turn into more work. Whoever curates that front page has a sense of humor.

The botsitting argument is simple and mostly right. You hand a task to an agent, then spend the next nineteen minutes watching it, correcting it, re-prompting it, and finally doing the thing yourself. Net gain: negative. The productivity claim was real, the labour that paid for it just never landed on anybody’s timesheet.

I buy the premise. The accounting is where I get off the bus.

What the stopwatch is actually catching

Take one of those nineteen-minute sessions and split the time by category instead of totalling it.

Some of it is genuine supervision. The agent grabbed the wrong row, misread a date, clicked into a settings page you now have to back out of. That’s the real tax, and it’s the one that only shrinks as models get better, which is slow and mostly outside your control.

But a large share of that time is not supervision. It’s reconstruction. You’re pasting the client name in again. You’re explaining that “the dashboard” means the Metabase instance and not the Stripe one. You’re solving a login you already solved on Tuesday, inside a browser that has never met you, for a session that will forget all of it in forty minutes anyway.

That second pile is not the price of AI being imperfect. It’s the price of an agent that starts from zero every single time you talk to it.

The blank tab problem

A cloud agent spins up a fresh Chromium somewhere in a datacenter. No cookies, no SSO, no idea which of your four Google accounts is the work one. So before it can do anything useful it needs you to hand over the entire world: credentials, or a screenshot, or three paragraphs describing a page you are currently staring at. Then a 2FA code arrives on your phone and you type it into a browser you cannot see, which is a genuinely stupid way to spend a Tuesday afternoon.

Every one of those minutes gets counted as botsitting. None of them are the model’s fault. They’re the plumbing’s.

Three things that stop being work

When the agent runs inside the Chrome you already have open, this is what falls off the clock:

Auth. There is nothing to log into. The tab is the session. Okta already waved you through this morning and the agent inherits that, the same way a second tab does.

Context pasting. No “here is the ticket, here is the customer, here is what I’m trying to do.” It’s reading the page you’re reading. The state is the prompt. I’ve written before about the context tax and this is the part of it that actually disappears rather than shrinking.

Handoff. The output lands in the form, not in a chat window you then copy out of. That last step sounds trivial and it’s the one I underestimated the longest.

Watching is not babysitting

Supervision only feels like unpaid labour when you can’t see what’s happening. That’s the distinction the HN post flattens.

An agent working in a side panel next to the page is doing something you can follow at a glance while you read a Slack thread. You see the click before the click, so a wrong turn costs two seconds and a stop button, instead of surfacing eight steps later as a form submitted to the wrong vendor. Compare that to a cloud run where you get a status spinner, then a summary, then the slow forensic work of figuring out which of the eleven steps went sideways. I’d rather watch than audit. The whole reason agents going off-script is scary is that most of them run where nobody’s looking.

Dassi sits in that panel because of this and not for any design reason. Same tab, same login, same screen. You supervise the way you’d supervise a colleague sharing their screen, which is to say: loosely, and only when something looks off.

The bit I can’t argue away

Models still get things wrong. Dassi included. That part of the tax is real and I’m not going to pretend a browser extension deletes it.

So what’s left

Strip out the re-authenticating, the re-pasting, the re-explaining, and what remains is correction. Actual correction, of actual mistakes, which is a cost you can reason about and which gets cheaper every time somebody ships a better model.

The hell of the current numbers is that nobody separates those two piles. The botsitting post counts them as one lump and concludes the whole category is a wash. I think it’s counting the plumbing and calling it the model.

Anyway. My agent is three tabs away filling in an expense form I have been avoiding since Thursday, and I have checked on it exactly once.


Written to `src/content/blog/botsitting-supervision-tax-browser-agents.md` (~840 body words, zero em dashes).