I’ve been doing quite a bit of agentic programming, which is comprised of my own Pi agent. But I am not the type to spin up 16 Claude Code instances and just say “make me a frontend for this.” That approach, to me, is just asking for trouble. You need real domain knowledge to tame these tools, without it, you are not directing an agent, you are just hoping you get some output, which to be fair, usually works with frontier models like Claude Opus 4.6 or Codex 5.3, but that’s still…not good.

Whenever someone asks me how to start programming, my answer is always the same: read a book, write code by hand, understand how specific concepts actually work. AI tools can accelerate learning, sure, but only if you already have enough foundation to tell when they are wrong.

Point being, knowledge has been more important than ever. Even if you don’t use agents (which by the way, completely OK and respectable!), you still are accountable for your code, and an agent writing your code doesn’t change that. I treat agents as linters or someone to prototype with, I still retain full control, ideas, critical thinking and the knowledge I have acquired over the years to build my ideas and bring them to life. There’s a lot of depth to understanding how to use the tools and plenty of traps to avoid.

Once again, human reinforcement is the way forward, keep learning, exploring, understanding and trying new things out. You should absolutely understand how the pieces fit together, how data in these systems work and fail so you will be able to make reliable systems, security, devops, research skills; much of which fill the whole puzzle.

Finally, interact with your fellow humans yourself!