The paradigm shift that we are in is now real. I just realized how much my own world has shifted in what and how I do things at work, in less than a year. December 2025 was the first time I allowed an agent to write a feature from a prompt and felt safe that the result would be good enough to use.
I’ve been thinking about welding and punch cards - and that has helped me think more freely about where we are heading. And what my value will be in that world.
Welding
Back in the 1960s–80s the industry was taken over by industrial robots. Jobs like welding, painting and assembly moved from being done by a human to being done by a robot.
Where a welder used to do the welding, a robot took over much of the work. The welder’s job was now moved a bit up in the value chain; to instruct (called programming) the robot on how and where to weld, but also to check the work after the robot was finished. Sometimes other machines (like x-rays etc.) were used to automate the checks and validate the work.
If a welding seam (this is the full extent of my welding knowledge being flexed here) was not good, the instructions to the robot were changed to ensure that the welding systematically became better. I’m sure they also redid some of the welding that was bad, but the important change was to make the system better.
This is VERY much like how the programmer’s job has moved from being a computer programmer to becoming a product developer. What I’m spending my time doing is instructing the agent, using prompts but even more with harnesses and skills. I also write a lot of deterministic validation code that uses other tools like automated tests, SonarQube checks and ESLint to check the code generated by the agents.
In the end I will need to validate that the functionality works as intended.
I have, like the welder in the 80s, moved up in the value chain and my effort goes into instructing the agent and validating the work that the agent does. I use many automated tools to do so and do some checks manually.
Punch cards
I’m old enough to have met people that worked with punch cards. Punch cards were a way to store data and programs. I remember the conversations with these old programmers:
- Marcus: Get out!?! Cards? As in real paper?
- Old programmer: Yeah, sure. We had boxes of boxes of them. That’s what those shelves over there was used for.
- M: But wasn’t it cumbersome?
- OP: You betcha. Once I dropped a box, and we spent hours putting the cards back in it. In the correct order.
- M: How on earth was it entered into the computer?
- OP: There was a reader and a card hopper or feeder. The reader sometimes got stuck due to dust accumulating from the edges of the cards as it was read. If one got stuck, you had to start over.
That feels like the stone age now, but in fact I know that it was used for some old untouched systems when I started university training in 1994.
Imagine the conversation that I will have with the young guns in just 5–10 years. They will ask me:
- Young gun: You WROTE the code? By hand?
- Marcus: Yes, line by line. We used something called text editors. People used to like one editor over another.
- YG: Wasn’t that hard?
- M: Hard?! Son, if I missed a semicolon on one line the whole program would not work.
- YG: How could you find that?
- M: We ran checks to ensure that the code was properly formatted. But that was the easy part - correctly formatted code could still mess up the memory of the program and crash the whole thing.
What we have done for a long time and spent millions to build tools and communities around (Svelte is better than React, right?) will soon feel antiquated, I think. Now I’ve used programming as an example, but I can totally see this in other areas too (No, you don't understand. I had to ensure that the boxes on the slides I created were properly aligned by squinting and dragging them. Yes! By hand.)
The future
In an office in 2036, the conversation with a young gun continues, after Marcus opened a few files and showed some code:
- Young gun: OMG - how would you know if this is good or not?
- Marcus: Well… I wrote this by hand for 20 years. I know and have a few heuristics. For example, no lines should be longer than 80 characters or they will not fit on my editor and then …
- YG: Ok - but how can I know? I have never written a line of code in my life, but I have instructed agents for 5 years.
- M: I don’t know. How could you know?
- YG: I don’t know.
- M: Me neither, little sister. Me neither. You will need to find your own way.
That conversation has no expert in it. Everyone alive today who can tell you whether agent-written code is any good learned it by writing code by hand, for years. There is nobody else to ask. That won’t be true forever, but it’s true now, and I notice that the people announcing loudest that nobody needs to code any more aren’t the ones asking.
Punching holes in the right place, welding perfect seams, writing semicolons at the end of each line - this was never the job. The job was to solve problems for the users of our product. Doing these things just happened to be the way we solved the business problem at that point in time.
But typing was also how we learned to see. The welder can judge a seam because he burnt his own thumbs making bad ones. Twenty years of my own bad code is where my heuristics come from. We’ve now automated the typing and the apprenticeship in one go, and only one of those was on purpose.
Our young gun above will find her own way, I’m sure of it. And when she does I am looking forward to learning from her, and the new ways that she has picked.
In Dec-25 I wrote the first feature from a prompt and didn’t have to touch it to make it work. Now, Sept-26, my agents are spinning up subagents that are communicating, criticizing each other’s work and reporting back to me. All of them follow my high-level instructions and are validated using a mix of LLM (non-deterministic) and deterministic validation tools.
We have just seen the start of this, I reckon.