The rise and fall of programmers

Right now an engineer paid six figures is waiting.

His agent is drafting a Slack reply to a message his coworker’s agent wrote, about a ticket neither of them has read.

He’ll send the draft without reading it. Then he’ll get a coffee and scroll X until the next prompt finishes.

If you write software for a living, you know this person. Maybe this week you were him.

Across town, someone with the same models and the same laptop shipped a product on Tuesday and got paid by strangers on Wednesday.

Same tools. Opposite years.

By the end you should be able to tell which way you’re drifting, and why the line isn’t where most people draw it.

The problems that paid the rent

For twenty years, most programming work was shallow in a very specific way. Build the CRUD app. Wire up login. Integrate Stripe. Write the scraper, the admin panel, the endpoint that moves JSON from here to there. Grind the LeetCode medium that gets you the job where you do all of the above.

Those problems were hard enough that you needed a professional, and easy enough that you could learn them from a book.

The second half is what killed them.

Stanford researchers, working with ADP payroll data, found a gap.

Employment for 22-to-25-year-olds in the most AI-exposed jobs is 19% below where it would be if it had kept pace with their less-exposed peers. Experienced workers show no such gap.

The split follows a line. Jobs built on codified knowledge, the kind written down in textbooks and docs, are shrinking. Jobs built on tacit knowledge, the kind you only get from practice and mistakes, are growing.

A model is made of text.

Whatever was written down, it already knows.

The shallow problems didn’t get harder. They got free.

The unemployed didn’t get worse. The ladder lost a rung.

Recent computer science graduates face 7% unemployment, against 4.2% for all recent graduates, according to the New York Fed. That ties CS with performing arts.

The easy story is that they’re lazy, or reading ChatGPT answers off a second screen during the Zoom interview. Some are. It’s one reason companies are dragging interviews back into the office.

The Stanford data tells a less comfortable story. The drop comes mostly from companies not hiring young people, not from firing them.

The shallow problems were never just the work. They were the training.

Juniors learned the deep stuff by doing the shallow stuff a thousand times, next to someone who already knew how.

Now the agent does it a thousand times, and nobody learns anything.

The employed aren’t winning either

Developers told BairesDev that AI now saves them 13 hours a week on coding, up from 7 a year ago. Then they said where the hours went. Two out of three spend more time reviewing AI-generated code. Half spend more time debugging problems the AI introduced.

The time saved never came back to them.

It went to reading what the machine wrote.

It isn’t only code. Stanford and BetterUp found that 40% of US desk workers had received “workslop” in the past month: AI output that looks finished and isn’t. Each piece took about two hours to clean up.

The engineer at the coffee machine is where this ends.

His agent writes. His coworker’s agent answers. Two people paid six figures act as routers between them.

Nobody is thinking. Everybody is busy.

That’s not a job being automated. It’s a job being hollowed out while the paycheck stays the same.

The average idea

The obvious escape is to quit and build something. Plenty of people try. They open a chat and type: give me ten SaaS ideas I can build with AI.

The model obliges. It hands them the same ten ideas it handed everyone else who asked that week. An AI wrapper for invoices. A habit tracker. A chatbot for dentists.

A model is the average of the internet.

Ask it for an idea and you get the average idea, already built by the hundred people who asked before you.

I built an agent whose only job is to find me business ideas. On its last run it found three, and its own red team killed all three. So far it’s much better at killing ideas than at finding them.

That’s the job it’s good for. The idea itself has to come from something you know that isn’t written down anywhere.

The ones who rose

Simon Willison, who builds open-source tools on his own, told Business Insider that coding agents made his work “a lot more fulfilling.” He now takes on projects that once needed a small team.

His explanation: “Using coding agents well is a deep skill,” one that takes both real engineering experience and experience with the agents, “and I have both.”

Stripe looked at the top 100 AI companies on its platform. The median one reached $1 million in annualized revenue in 11.5 months. Four months faster than the fastest SaaS companies of the previous generation.

They didn’t spend the saved hours reviewing more code.

They spent them one layer up. On a customer. On taste. On distribution. On problems that don’t fit in a prompt, because nobody has written them down yet.

When a problem becomes free to solve, the value moves to the problem next to it.

The shippers moved. The relays stayed at the old problem, getting paid to watch it solve itself.

Rise or fall

Programming isn’t dying. The shallow version is.

For a lot of us, the shallow version was the whole job.

The fall is handing the agent the work and keeping none of the judgment.

The rise is using the agent to clear the easy problems so you can reach the hard ones. The customer. The market. The thing only you have seen.

So the next time your agent finishes a task, ask yourself:

What’s left that only you could have done?

What’s left is the part that was never written down.

The customer you actually talked to. The mistake you recognize because you made the last one. The idea an agent can argue against and still can’t invent.

If the task finishes and nothing remains that you decided, you didn’t get the hours back.

You became the person at the coffee machine.