*A version of this was presented at *the PugTO (programmer user group meetup) in Italy in May 2026 .
A professional footballer does not play until retirement. At some point the body stops, the role changes, the career evolves. It seemed obvious to me when I started thinking about it and then I realized I had never applied the same logic to myself. Why? Because we have spent years fighting against it discussing about this with posts like “The Engineer/Manager Pendulum” from Charity Majors, particularly in Italy where the Staff/Principal career is still today a snowflake.
With AI it looks like on average the skills per head is shrinking down, middle manager are being layoff pretty much everywhere, many developers are fine turning to part time product people as well. I am don’t think will stay as it for long but there is for sure an attempt to reshuffle competence.
Sean Goedecke puts it plainly: until around 2024, the best way to learn software engineering was just doing software engineering. That was lucky for us. It meant a coding hobby could become a lucrative career, and that the people who really liked the work just got better over time. But that was never an immutable fact about what software engineering is. It was a fortunate coincidence.
I think the coincidence might be ending.
I don’t think there’s compelling evidence that using AI makes you less intelligent overall. However, it seems pretty obvious that using AI to perform a task means you don’t learn as much about performing that task.
That quote has been sitting with me since I read it. Not because it is alarming it is just accurate. And then Goedecke makes the turn that is harder to argue with: we might still be obliged to use AI even if it atrophies our skills, for the same reason construction workers are obliged to lift heavy objects. Because that is what we are being paid to do. Construction workers do not say that being good at their job means not lifting heavy objects. They say: too bad, that is the job.
I have been thinking about this while doing my actual work. These days I often find myself in what I can only call yes-man mode: my main contribution is telling Claude it is doing well and to keep going. It makes me smile. It also makes me think.
Addy Osmani, Director at Google Cloud AI, names two things worth keeping separate. Cognitive death is when you stop understanding what is happening in your system. Cognitive surrender is when you blindly accept what the AI gives you. The first is a slow erosion. The second is a choice. Both compound quietly.
Margaret-Anne Storey calls this cognitive debt — the gap between a system’s evolving structure and the team’s shared understanding of why it works. When you use AI to move fast without building that understanding, the debt accumulates. Not in the code. In the people.
The risk here is not hypothetical. Mark Seemann adds an economic dimension that most people skip: the systems we are all depending on right now are running at colossal deficits. OpenAI and Anthropic are losing money every year. The way agentic software development works today is not representative of future directions. We are training our habits on subsidized tools whose economics are not yet resolved.
There is a different way to use these tools, though. Nolan Lawson writes about using AI to write better code, more slowly. Not to ship more, but to think more carefully — generating tests, exploring edge cases, examining assumptions. Speed as a choice rather than a default.
That framing helps me. I am spending more time thinking about what I actually like doing and what I do not. Not because AI is taking over, but because the question of what the job is has genuinely opened up in a way it was not before.
I presented a version of this at a meetup in Italy recently. The room was full of people who had been writing software for years. Nobody had clean answers. But the question that stayed with me afterward was this: if our superpower was always the ability to talk to machines — to translate between human intent and machine execution — and that translation is becoming easier for everyone, what was the actual value we were providing?
I do not think talking to machines was ever the real superpower. I think understanding the problem well enough to know what to ask for was. That part is harder to automate.
That is what I am trying to hold onto.
This meetup helped to reinforce something I experienced quite often those days. Agents are helpful but the cognitive load to make them to do what you want, how you want it and to trust the outcome in a way that makes you confident in maintaining it long term is pretty high, so much that often the feeling is that writing the core long term maintainable version of a feature is still cheaper to do manually. Agents can help to build ten, hundreds of versions of it, to gain confidence of design, try different approach, but the long term one, not sure yet.
This is not enough to skip this technology, it is revolutionizing how we write software and we should learn how to use it. I saw many amazing developers who wrote code we use everyday skipping autocompletes capabilities or IDE, but the same people today are using LLMs.
Practical lessons on shipping software, straight to your inbox. No fluff.