Coding agents are GREAT until they AREN'T,

Everything started on a Saturday night with me trying this shiny new object called Cursor (back before the pricing changes). I tried a few prompts, and I was shocked that it could actually write code. It was great.

Coincidentally, the week after that, a friend contacted me and said,

“I'll need to do some kind of quiz, etc… But Google Forms isn’t enough. Can you make that in FOUR DAYS?”

The most appropriate answer would have been a simple “NO.” But this time was different; this time I had super‑powers.

So I grabbed my coffee and started programming with Cursor. I was the architect, and it was the coder. I was telling it do this and that. When it got stucked I helped it. And yes, finally, IN 4 DAYS we had the product ready. It was used on Friday by an audience of 100 people (I know, I know— not impressive, but they were all using it within the same 15‑minute window, and there was a real‑time leaderboard, etc.). It simply WORKED. So I felt an even stronger urge to keep using it.

Then Cursor changed its pricing…

After the pricing change, I started looking for other solutions. Gemini CLI—wow, it was free! But it sucked. I was addicted; I wanted to command something to build applications for me. I ended up subscribing to Claude. And what a tool it was! Claude 4 Sonnet was simply amazing. Most of the time, I wasn’t even checking the code; it was just working like magic.

So I thought, Maybe I can just single‑shot an entire product. I trusted Claude Code that much. I wrote some scripts and managed to run Claude for literally hours. But the result… was terrible.

I had developed an internal application before with vibe coding, and I needed to make critical changes to improve the overall experience and performance of the app. Instead of writing it myself like a human being, I ordered Claude Code to do it and waited for the results. It created so many functions and pages, etc. At first glance, it seemed legit. It even “optimized” everything according to the device’s RAM. There was only one problem: even the prior functionality was gone, and nothing was working.

So I decided to get my hands dirty. For the first time ever, I checked the code—and what a mess it was! No proper app state, duplicate functions everywhere, and the RAM optimization I mentioned was literally:

// TODO: this will be implemented
return dummyData;

So, if you’re a developer, you’re still safe—at least until the next model release. I truly believe that, when used properly (and without leaving all the thinking to the model), AI increases productivity. But don’t let it decide everything—at least for now!

Thanks for reading. Oh, and invest your money wisely while you still have a job. 🙂

This blog post was written with ❤️ and lots of ☕