My Agents Make the Mistakes

31 Jul 2026

Nowadays, I feel a bit lost and conflicted. It feels like I have stopped developing and become someone who is no longer truly learning, but only trying to memorize things.

I use AI a lot, almost for everything. I use it to write code, discuss ideas, and get information that I would previously have spent much more time finding on my own.

The other day, I started thinking about what used to make me happy in this job before AI became such a big part of it. I realized that there were three main things:

• The feeling of learning and making progress

• Being recognized by others

• Getting things done

If we give these three things equal importance, coding agents mostly provide only the third one. That means my satisfaction has dropped to around 1/3 of what it used to be.

But the reality feels even worse than that.

Nowadays, almost everyone can get things done. We have very intelligent and capable models, so simply producing a working result no longer feels as valuable as it once did. That final 1/3 sometimes feels more like 1/12.

The interesting part is that LLMs are still mostly brute-forcing their way through problems. Because of that, if you have deep knowledge of your field and can reason carefully about your work, you can still create very strong results, sometimes even state-of-the-art results. These could be compilers, fast kernels, good system designs, or many other things.

However, another problem starts here.

Everyone is in a rush because it has become so easy to make something work. If a solution works 70 percent of the time, the final 30 percent is often ignored because the result is already better than the having nothing. This creates a deeper problem: a shift in the way we build software.

Until now, we have designed systems to be readable, maintainable, and easy to extend. But LLMs often work differently. They tend to generate scripts or solutions built for one specific task. When a new requirement appears, instead of extending the existing system, it can feel easier to generate another solution from scratch because it is cheap and fast.

If you are a person like me, you may find yourself spending most of your time trying to understand what the agent did, why it did it that way, and how it could be improved. At first, I thought this was a good way to learn faster and design better. After using this approach for a long time, I now think it was a bad idea. I realized that I was not really learning. I was mostly memorizing the outputs.

Real learning comes from making mistakes, trying different approaches, reading, failing, and trying again. If you are not the one making the mistakes, you may not be the one learning from them. My agents make the mistakes, while I am left trying to understand the final result. This revolution is going to happen or happening whether we like it or not, and I am still trying to find the best setup for myself.

So far, I have learned that I need to spend more time reasoning about what I want to do and why I want to do it. I also need to take smaller steps instead of offloading large parts of the work to agents. I want to move step by step and properly understand what I write, test, and experiment with. Honestly, I still do not know what the best approach is. I also do not know how to stay involved and continue developing my skills while agentic LLMs getting intelligent so quickly. But I think the first step is to stop treating speed as the only goal and start protecting the parts of the work that help me learn, think, and grow.

Written by human, polished by a LLM