I use AI coding agents every working day, and they have made me faster. I also once watched an AI assistant make me a worse engineer inside a single week — and I felt it happen. Both things are true. Pretending otherwise, in either direction, produces the useless takes: “AI rots your brain” from one side, “skill worry is just fear” from the other. This is my attempt at the honest middle, from the inside.
What gets weaker, exactly
The week that scared me, I was running an aggressive autocomplete assistant. I noticed I had stopped finishing my own thoughts. I would type half an idea and wait for the machine to complete it — then judge its suggestion instead of forming my own. The muscle that produces a solution from nothing was resting, and a resting muscle gets weaker. When I turned the assistant off for an afternoon, the first hour was slow in a way that embarrassed me.
What fades is not intelligence. It is specific, quiet abilities. Holding a problem in your head long enough to feel its shape. The debugging instinct that says “the bug is not where the error is.” The patience for the uncomfortable minutes of not knowing — which is where real understanding gets built. None of these have a syntax. All of them shrink when every uncomfortable minute is interrupted by a suggestion.
What gets sharper
And still — other abilities have grown. Working with agents every day forces me to state my intent with a precision I never needed before. A vague thought now produces vague code, instantly, at scale, so vagueness finally has a visible cost. Reviewing machine output all day has made me a sharper reader of code than writing it ever did. And the expensive skill has moved up one level: from “how do I implement this?” to “what should exist, and how will I know it is correct?” That is not a loss. That is the job getting better.
The dependency test
My whole career philosophy is empowerment over dependency. I spent years teaching clients to run their own websites instead of renting my hands forever. Turning that lens on myself was uncomfortable: is the tool empowering me, or am I becoming its client? My test is simple. Could I still do this without the tool — slower, but properly? If yes, the tool is leverage. If no, and the skill is one I decided matters, then I have outsourced something I still needed to own — silently, to a service I do not control.
The words “I decided” carry the weight there. I let mental arithmetic fade years ago, like everyone with a calculator, and I do not miss it. The failure is not letting skills go. The failure is letting the tool choose which ones, without noticing the trade.
Habits that survived
- Read the diff before the explanation. Form your own judgment of the code, then read the agent’s story about it. In the other order, you are not reviewing — you are being persuaded.
- Predict before you delegate. Before handing a problem to the agent, I spend one minute writing what I think the answer is. It keeps the generation muscle working, and it turns every delegation into a calibration exercise — I learn exactly where my model of the system is wrong.
- Keep manual repetitions, on purpose. Some problems I solve without help, chosen deliberately. It is slower. That is the point.
- Protect the not-knowing minutes. When a problem is interesting, I sit with it before asking. The discomfort is the exercise. The research on all of this is still young — which is exactly why I would rather build the habits now than wait for the studies to tell me what I lost.
The tools are staying, and I want them to — they build most of my software now. But thinking is a use-it-or-lose-it system, and the connection between my mind and these machines is now the most important system I maintain. I plan to maintain it the way I maintain production: with monitoring, with named failure modes, and with a deliberate fire drill from time to time.