The Great Suspense

Do you know the cartoon trope where a character runs off of a cliff, but does not fall down immediately? He’s just there, hovering in the air — or, as the laws of cartoon physics put it — suspended in space.

This is how I feel about AI coding right now.

I distinctly remember the time a colleague first showed us how an AI agent can just take his loosely phrased change request and update the code accordingly. I was impressed, but I wasn’t enthusiastic (read my other post to understand why).

But above all, I was skeptical.

I saw my colleague look at the diff the AI recommended and accept it, and I thought to myself: that’s all nice and well, but the only reason you can decide if the AI produced correct code is because you have extensive experience in doing this yourself, ie. coding by hand. If you start relying on the AI to produce all your code, you’ll lose the ability to make an informed judgement about what it has produced.

Well, after staying away from agentic AI for longer than most of my teammates, I have been increasingly relying on it to produce code for me too. This is what makes me feel like having run off a cliff in terms of knowing how to code, and telling good code apart from bad code. The momentum of is still driving me forward but the ground has vanished from under my feet. And so soon enough, I’ll start falling into the abyss of not knowing how to code anymore. I’ll only know how to work with the agent.

And so the million-dollar question is: does it matter?
Personal feelings aside, is it a problem for my work, for all of our work, if we can’t code anymore?

Because that is one of the main line of argumentation against surrendering to the AI wave: the statement that

• when the AI can’t fix a bug, you’ll need a human to dig into the codebase and resolve the issue;
• when the codebase becomes a mess, you’ll need a human to refactor it back to a usable state;
• when there is an outage, you’ll need a human to undertand what went wrong and how to fix it.

But is this really the case? And if it is, will it stay like this?

Will human coding expertise still be needed, whether that’s due to the inherent limitations of large language models, unsustainable pricing of frontier models, or any other reason? Or will the improving models and harnesses cause us to say goodbye to writing code by hand for good?

In a workplace conversation about AI half a year or so ago, I expressed my hope for more clarity and certainty to come with time. The desire to see more clearly where we really stand with agentic AI and the hope that we’ll get that soon enough. Not knowing what will happen is the other great suspense that is stressing me nowadays.

Am I bound to crash into ground at the bottom of the canyon, or will I find that where we’re going, we don’t need roads?