Developing with Agents

The other day, I launched an updated version of my website, which was pretty much completely rebuilt from scratch on a new CMS (Grav) and redesigned to match modern standards (with all the effects, cool features, responsive design, and all that jazz). And here’s what I want to share: this was probably my first full-scale, hands-on experience with serious development assisted by neural networks (which people for some reason call "AI").

By "full-scale experience," I mean something way more complex than a basic Python script or just asking Gemini for a snippet of code. I'm talking about working with agents in VSCode, Docker, GitHub, and the whole nine yards. What makes this interesting is that, as someone who has been writing code in various languages for over 20 years (though I am by no means a professional programmer) and is generally familiar with web dev, I actually had a solid baseline to compare this "vibe-coding" process to.

And man, it's a fascinating thing. What blows my mind the most is how much this technology "democratizes" development, especially for personal projects. I might be out of touch with current market rates for this kind of web development, but either way, I don't just have two or three thousand dollars lying around to pay a decent developer. I also don't have the time (nor desire) to master the whole tech stack (Grav CMS, Tailwind CSS, a ton of specific JS libraries) just to build a personal portfolio. But I still wanted a cool, original website :) And neural networks made it possible—not in 3 to 4 months, but in just a couple of weeks of evening sessions!

On top of that, what really wins you over is how incredibly... comfortable the whole process feels. I’m not giving formalized, rigid instructions to a "robot." I’m chatting with the LLM in my native language like an old friend—complete with all my slang, jokes, and idioms—and it gets it perfectly. Well... except for those occasional hellish brain farts at the end of a multi-hour session, when the agent’s context window just gets completely clogged up and you have to give it a hard shake to snap it out of it.

I think it's important to make one crucial disclaimer here: throughout the development process, there were plenty of times when my own background and understanding of the underlying processes were absolutely necessary, despite how smart the LLM was. In other words, finding a bug in the code and fixing it with my own hands was often times faster (and sometimes the only viable option) than endlessly jerking off stuck agent.

Long story short, in my humble opinion, AI agents are an awesome tool. They don't replace a specialist, but when approached the right way, they make you wildly more efficient!