Amazon quietly stopped accepting new customers for Mechanical Turk, and I keep thinking about what that marketplace was actually for. Not the grand “artificial artificial intelligence” pitch Jeff Bezos gave it back in 2005. The real work. Someone had a spreadsheet of 4,000 product URLs and needed a human to open each one, find the price, and type it into a cell. That was MTurk. A firehose of tiny, boring, browser-shaped tasks that no software could reliably do, so you rented strangers to do them a few cents at a time.

That firehose is drying up. And the reason isn’t that the tasks stopped existing.

The work never went away, the crowd did

Go read a random sample of old MTurk HITs and it’s almost funny how specific they are. Transcribe the phone number from this business listing. Categorize whether this tweet is positive or negative. Copy the three bullet points from this webpage into a form. Find the CEO’s name on this company’s About page.

Every one of those is a thing you do inside a browser tab. Open page, read page, extract fact, put fact somewhere. MTurk existed because computers were bad at the “read page” step when the page was messy, inconsistent, and built for human eyes instead of an API. So Amazon built a labor exchange to route that step to actual eyes.

The models got good at the read-page step. That’s the whole story, really. What used to require a distributed crowd of people squinting at listings now happens in one pass by a language model that can look at a rendered page and tell you the price, the phone number, the sentiment, the CEO’s name. The demand didn’t collapse. The supply side just moved from humans to software.

What a requester used to fight

Here’s a detail that gets lost. Setting up an MTurk job was its own small hell. You wrote instructions clear enough for a stranger to follow, you built approval logic, you handled the workers who rushed through and got it wrong, you dealt with tasks that needed a login the worker didn’t have. If the data lived behind your own authenticated dashboard, MTurk was basically useless, because you couldn’t exactly mail your session cookie to a thousand contractors in Hyderabad.

So a huge category of grunt work was never even eligible for the crowd. Anything sitting behind your own login stayed manual. You did it yourself, or you paid an offshore VA and shipped them credentials over Slack and prayed.

That’s the part I find genuinely interesting, because it’s exactly the gap a browser agent closes. The agent isn’t a stranger you’re handing data to. It’s running in the tab you already opened, using the session you already authenticated, seeing the page you already have permission to see. No OAuth handoff, no cookie sharing, no “please screenshot the invoice and email it to the contractor.” This is roughly the argument we made in Browser Agents Are the New API, and MTurk shutting its doors is a weirdly precise confirmation of it.

A concrete swap

Say you’ve got 60 orders sitting in your Shopify admin and you need each customer’s city pulled into a sheet for a shipping estimate. Five years ago that’s a HIT, or an afternoon of your own clicking.

Now you open the tab, open Dassi in the side panel, and say: go through these orders, grab the shipping city for each one, and build me a table. It reads what you’re looking at. It clicks through. It hands you the table. The data never left your logged-in session, never got mailed to a worker, never touched a third-party server it didn’t need to. The thing that made this task un-crowdsourceable, the login, is the exact thing that makes it trivial for an agent in your browser.

You can point Dassi at the same categorize-transcribe-extract chores MTurk was built to farm out, except it works on the pages only you can reach.

Cheaper than a nickel a task

MTurk’s economics were always a little grim. Pennies per task, and even at pennies the numbers added up when you had tens of thousands of rows. People built entire businesses on arbitraging that gap between “too small to automate” and “too big to do myself.”

An agent collapses that gap. Because the marginal cost of asking it to extract one more field, or a hundred more, rounds to nothing, especially if you’re on BYOK and paying your LLM provider directly instead of a marketplace taking its cut. The tasks that were worth exactly three cents to route to a human are now worth zero cents to route to the tab.

I don’t think Amazon killed Mechanical Turk out of some grand AI strategy. Marketplaces just die when both sides quietly leave. Requesters stopped posting the easy stuff because their own tools started doing it, and once the easy volume goes, the whole thing gets thin. Twenty years is a good run for anything on the internet. The microtasks outlived the marketplace, which is a strange kind of legacy. They just live in your tab now, and nobody’s getting paid a nickel to click.