I have a confession to make to some of my creative friends - I have used AI. Specifically agentic AI, which is the kind where you enter a prompt and it explores your codebase and does what you ask of it.

At first this felt like a great way to catch up on the many side projects that I haven't finished. I explored different kinds of agentic AI, such as Mistral Vibe, Pi.Dev with OpenRouter, and Jetbrains Junie, which lets you choose different models, based on cost. I steer clear of Claude and OpenAI as I do not wish to support these companies.

I totally fell into the trap of prompting to develop things, instead of actually writing the code myself. I know many in the creative X tech community (eg games) who steer clear of all AI, autocomplete, everything. Which I totally respect and think highly of those who do that.

The very simple lesson I learned with agentic AI was that it created code I didn't like. I kept adding features to things, and the code it generated was a bit of a mess, with code in the wrong places, poorly split up, despite telling AGENTS.md to reuse code wherever possible. I learned enough about agentic AI that you can steer it in the right direction, but I argue that it's not enough to create maintainable code. A few that I know have predicted this before I experimentally discovered it. Which makes me feel a bit dumb for not listening. Maybe it is a skill issue, but who is teaching agentic coding? Certainly not the AI providers themselves.

With agentic coding getting "better", it worries me that people will let go of coding all together, not even looking at the code that is generated. This was the concern initially, but hasn't gone away. The real concern is major software companies, and of course open source projects on GitHub that provide software that people use, is being made by machines and not thoroughly checked by humans. My prediction if things continue this way, is that programming languages themselves will have to change into something that the machines prefer, more succinct, perhaps not based in English. Then, humans won't be able to maintain it as well. This is not a good outcome - if we all give in to the vibe coding temptation.

I have a historic issue of difficulty maintaining focus during the day, and AI certainly appeals as a cure for my limited work capacity. But less so agentic coding. I have proven to myself that I can write code fast, solve problems fast, but since I multi task so much, it's still easier to prompt Junie/etc while I work on something else. But, my thinking is, if I choose to use AI to solve a problem, is AI really necessary? Can I not either look for a simpler solution, or just manage my time better. On a mental health and wellness perspective, there are things I can do to improve my concentration, like walks, meditation, diet, etc. Instead of vibe coding I can go for a walk to clear my head.

So, for now I plan to go back to writing virtually all of my main projects by hand. With a little AI assistance when I get stuck, but it is a back and forward thing. I'll use it and go, do'h of course that's how to solve that problem, then use AI less.

Agentic coding I'm not dropping altogether, as it is useful for rapid development of ideas that I simply do not have time for. These experiments won't go out to the public, and if I intend to keep working on them, I'll start fresh with my own codebase.

What using AI has taught me is to enjoy programming again. If I don't enjoy it, I should be doing something that I do enjoy. Is that the purpose of agentic coding? To replace coding that you don't enjoy? Some suggest it's to turn burned out programmers into managers, but idk.

If I am for some reason required to use agentic coding for work, I will try to insist not Claude or OpenAI. Unless they pay for the subscription, and all the tokens. I'm not strictly anti-AI as a result, but I am more determined than ever, even before generative AI to make the most of my own brain.