Skip to Content
Hesham Hussen
The friction between AI and software development

The friction between AI and software development

Agentic-first development elevated the engineering role into what it was always supposed to be. But a quieter side effect bothers me: skill atrophy, and the loss of the knowledge sharing that keeps a team fluent.

3 min read•

Since the start of this year, specifically since Opus 4.5, my workflow has become agentic-first. I barely write code myself anymore. I write specs and plans, let the model build, and review at the end.

Opus 4.5 was the first model, in my opinion, that could actually do feature development end-to-end without major handholding or hallucinations, provided you spec the work rigorously and break it into chunks that don't overwhelm it.

To be honest, for a while I lived in near-panic. The jump was beyond what I expected. I started thinking about alternative careers, something I suspect a lot of us have done at some point by now.

That phase didn't last. I quickly realized that writing code was never my whole job. A huge part of my time goes to meetings with stakeholders, refining vague topics, pushing back on nonsensical requirements, and getting everyone aligned in a messy context. That is the hard work that I'm not convinced AI can do it, at least until now :D

So in a way, AI has elevated the role into what was always expected of good engineers: a technical partner who thinks alongside Product and management about how an idea gets realized in an engineering-sound way and not just bolted on or vibe-coded on the fly.

Most people would agree with that framing. But lately a side effect has been bothering me, and I don't have an answer for it: skill atrophy, and the loss of knowledge sharing.

I earn my seat because of skills I built over ten years of hitting my head against walls while reading and writing code. If I stop doing that, where are those skills in five years? I don't know. But I doubt it's a good place.

The second part is subtler. When you review a colleague's PR, you're not just only checking compliance, you're absorbing what's happening in the codebase. Even on features you never touched, review keeps the whole team fluent. Outsource that to AI and the fluency quietly disappears.

Picture it two or three years out: someone in a meeting asks about a product feature. The engineer who "owns" it never really read it, the AI wrote it.

"Let me go ask Claude." At which point, why are we in the room at all? Just put Claude in the meeting.

I don't think the answer is to stop using these tools, that ship has sailed, and they genuinely make some things better I think. But I think we're going to have to be deliberate about what we choose to keep doing by hand, precisely because we no longer have to.