The AI Code Agent Hype: Real Progress vs. Expectations | Jawuil Pineda<br>Hire Me<br>🇪🇸 Español

Back to blog<br>How about we talk a bit about the general state of AI? Has it really advanced as much as everyone says?

Honestly, I feel it has, but not nearly as much as the hype makes it seem. At least when it comes to coding, the reality is quite different.

Last year, I started heavily using AI-assisted coding tools. I began the way most people probably did: asking ChatGPT, Claude, or DeepSeek for code (by the way, I was using DeepSeek way before it went viral in 2025), and then copying, pasting, and fixing a lot of errors.

Later, I started testing more integrated tools. I used Copilot (which I remember being pretty bad back then), moved on to Cursor, and also tried Windsurf, back when the free tiers were much more generous than they are now. I also discovered Augment Code , a VS Code extension that, for me, was the absolute best in 2025. Honestly, it was far better than Cursor and Copilot because the way it parsed the context of the entire local codebase and provided ultra-precise inline edits was next level. Unfortunately, they dropped their free plan, and since I couldn't afford subscriptions at the time, I worked with them as far as the free tier allowed and then did the rest by hand.

I feel like for most of that year, progress was mostly stagnant. For me, the real breakthrough came around October and November 2025. Around that time, you could feel that these agents were working significantly better. Cursor and Windsurf matured, and suddenly tools like OpenCode, Gemini CLI, and Antigravity dropped. It was incredible stuff—they provided rich context and went from just autocompleting lines to tackling complex tasks across multiple files. I couldn’t test Claude Code because it wasn't available in my country, but what I already had access to was amazing.

Late 2025 was when I truly felt the generational leap in coding agents. However, as of today (mid-2026), I honestly don't feel they've pushed much further beyond that point. Sure, there are speed improvements and newer model initializations, but we haven't seen the same disruptive impact we felt late last year. We'll have to wait and see.

Over the last few months, I’ve been trying to understand a bit more about how these agents actually work under the hood, since I'm building my own for a few experiments. Honestly, they are quite complex systems, and the more you study them, the more you realize it's not some "weird magic"—it's pure engineering. In fact, the deeper you look into it, the more you realize how difficult it actually is for it to replace you. Still, the progress is awesome. Let's see where this takes us.

Back to blog