AI Isn’t Just Answering Anymore. It’s Doing the Work.

For a couple of years I treated AI like a very fast librarian. You ask, it fetches. That picture broke in March 2023, and it broke in public. A developer named Toran Bruce Richards released a free program called AutoGPT, open for anyone to download, with a pitch that sounded ridiculous at the time. It came with no manual. You typed a goal in plain English, and the program went after it on its own.

It started as a mess

It took off fast, ranking within weeks among the most starred AI projects on GitHub, and passing 100,000 stars within months. And it was a glorious mess. It got trapped in loops, repeating the same failed step. It invented facts out of thin air. It chewed through paid API credits at an alarming rate, and nobody claimed any of it was reliable.

Still, it proved the point that mattered. What it showed was hard to miss: put a language model in charge of real tools, give it enough patience, and multi-step work gets done without a human holding its hand. That was the moment the chatbot era started turning into something else: software that moves on its own.

Then the big companies arrived

Then, on October 22, 2024, Anthropic put out a public beta of something new: "computer use" for its Claude 3.5 Sonnet model. It was the first time a leading AI model could see a computer screen and work the mouse and keyboard the way someone sitting at the desk would.

OpenAI answered on January 23, 2025, with Operator. It launched as a research preview for Pro users in the US, an agent with its own browser built on a single promise: tell it what you need, and it handles the repetitive clicking that eats up everyone's afternoon. Forms. Groceries. Both companies said the same two things at launch. Big step forward. Still makes mistakes.

What they actually do all day

Forget the demos. In practice, agents are good at exactly the work nobody wants. Research is the standout. Give it a research topic and it will go through dozens of pages before coming back with a structured report. Then there's the dull middle of office life. Data that needs moving between apps that refuse to talk to each other. The same forms, filled out every single week. Replies that need drafting. Files that need tidying. They write and test simple code too. Nothing magical. Just the same loop, pointed at jobs that are mostly reading and clicking.

The ten-step mistake

Here's what keeps me cautious. A chatbot's mistake costs ten seconds. Wrong answer, ask again, done. An agent's mistake can cost ten steps. It misunderstands you at step two. Then it keeps going, totally confident, and books the wrong thing. Emails the wrong person. Deletes the wrong file. You find out at the end. Mistakes don't just happen with agents. They snowball, at machine speed, while you're making coffee. That's the real danger, and nobody has fully solved it.

Your job is next on the list

If you read my piece about AI taking pieces of jobs, this is the sequel. Chatbots came for the answering. Agents are coming for the doing. First in line: the multi-step computer work nobody enjoys. Copying information between systems. Scheduling. Chasing replies. Assembling reports. If your day is the same mouse clicks in the same order, the agent isn't coming for you. It's coming for the sequence. The people who stay valuable will be the ones doing the things machines still fumble, like figuring out what should be done in the first place and owning it when things go wrong.

What agents can't do

They can't be accountable, for one thing. Book the wrong flight through an agent and the blame still lands on you, since you gave the order. They also have no sense of when a task is beyond them. A good employee stops and asks. An agent keeps clicking. So the human job changes shape. Less doing, more supervising. You draw the lines. You check the output. You shut it down early when something smells off.

Dullness advice that actually works

Here's my suggestion, and it's not exciting: give an agent one small job where a mistake costs you nothing, like sorting your downloads folder, and watch it work. You'll learn quickly where it shines and where it confidently does something stupid. Keep it away from your passwords, your money, and anything you can't undo until it's earned that access ten times over. Check its work like a manager, not a fan.

The people who'll get burned aren't the ones using agents. They're the ones trusting them.