Agentic Coding is a Terrible Experience

Programming VSCode Coding AI Agentic

AI in coding isn't what it seems. After numerous times trying, it’s like babysitting. Here's why I hate it.

Published on November 14, 2025

Not only are you basically not learning much of anything. You just end up a glorified babysitter making sure the LLM isn't going to destroy your entire machine with one single command, so you need to sit, endlessly allowing it to run commands (if you auto approve every command you are an idiot).

LLMs have basically taken over the coding world, and there are two camps. Those that proclaim they were able to create GTA 6 in 3 hours with Cursor, and those (me) who have actually used them, and know the limitations of them.

Full transparency, a mate of mine came to me with a business idea and fed it through AI, wanting to know how easy it would be to build. Now I've no real issue with this idea, so I said to him, give me the weekend, I'll throw my AI at it and see what it can do.

Knowing full well it was going to be a shit show.

I'm sitting here, just abjectly not caring about any of this. None of this is my code. And all the errors I'm seeing are just making me mad at an invisible coder failing every single time to do the bare minimum. I wanted a pretty simple test setup to the dashboard route, and it installed an entire end to end browser testing suite. Surely, for the love of god that isn't how Next.js does it's testing?!?!

This is a small snippet of this hour's long chat. We make some progress, do a thing, an error happens, it then spends the next 45 minutes going through each error one by one, having no clue what it's doing but the last thing it just did. It's an entirely dumb system, and I'd wager I'd have this working by now if I never used agentic AI at all.

And no, for the love of god, your model won't fix this, the next model won't fix it either. This is a core problem that needs owning up to. This isn't AI, it's a glorified auto complete that guesses what text character comes next. There isn't anything intelligent, it requires a dataset telling what to do in every scenario (hence why the models take so long to build). Hence, why you can't get it to count the amount of letters in a word. IT DOESN'T KNOW.

The best it can do, is take code from someone else, ideas from someone else, repackage them into a copyright free expression that governments seemingly the world over are just letting them do. "I didn't steal this from X company, I just got the AI to do it!" what a get out of lawsuit card.

I've decided. I hate AI. Furthermore, I'm leaving it at most for tab auto completes and doing everything my way. That way I'm bloody invested in the projects I'm making instead of sitting here like a drooling idiot waiting for magic to happen.

It ain't happening.

And yes, feel free to write me an email in 2 years if I'm wrong. I'm still waiting for GPT-4 to make me unemployed.

PS: This video from Primeagen is a must watch...

https://www.youtube.com/watch?v=AyuMdNoL1Vs

Related Posts

Gemini 3 Released, and it still sucks
AI Agentic +3 more

Gemini 3 Released, and it still sucks

I'm going to be extremely quick with this. Gemini 3 sucks, and failed at the first question I asked it.

Read More →
Laravel Reverb on Cloudflare on Forge
Programming Laravel +2 more

Laravel Reverb on Cloudflare on Forge

A setup from hell, it seems nobody has bothered to test this, the good thing is I've tested it so you don't have too.

Read More →