I WENT DEEP

Something is happening to devs, well, maybe to me only. I have coined it as agentic depression. Currently, agents have gotten better at code as compared to the early days. Not all, but the bleeding edge ones like opus, gpt5, gemini and kimi. Their wrappers are even shinier than ever, like Claude Code, Hermes agent, codex and kirocode to name a few.

I have been using agents to code for the past few months, opening up Neovim to light up the code here and there. But later it was full-on agentic drive in the later weeks. I simply compensated and fixed the code quality by spending every Friday cleaning up the code with the agent.

I went further with the agents by exploring Pi-coding-agents and building my own. Understanding how to build a knowledge base and all the exploration. It was fairly entertaining watching the AI engineer talk on YouTube, building my .agents/skills just right for every project. I learnt about compaction and context window management. My curiosity was all over the place.

I MISS THE OLD ABSTRACTION

The more I used my agent, the less out of touch I was with my code. But this is the abstraction. The ideas preached by the almighty Rich hickey. This abstraction is different from other abstractions, to be honest. I will compare AI abstraction with fennel abstraction. Fennel is a lisp dialect that compiles to Lua. Agentic programming is English that spews to code. The difference with the two is that in one, you are still in your creating bubble place, but instead of writing “boring” Lua, you write with what you are comfortable with, lisp. This kind of abstraction is “abstraction of comfort”

Same thing with using the reduce function in JavaScript or Clojure. Instead of spending time working on building your own “iterative function”, you use an already existing one. This kind of abstraction is “abstraction of effort”.

AI is argued to be powerful for those who know “what to do,” but it takes that out, the idea of knowing. Well, it is argued that if you prompt it right, you will learn something new. This is argued in many articles like the “don’t outsource learning”. One thing that is missed with such statements is that learning does not work like that. Learning is through having skin in the game; it is through combating a problem head-on.

AI can “teach” you, but as long as there is zero effort in what you are doing, knowledge will fly out quicker than it came.

WHAT ARE YOU DOING

What do you do when your agent is working? That’s the question one of the core team members of opencode asked on Twitter, and no one had a solid answer. I definitely do not have an answer. I have tried multiple things like reading the thinking output, reading a book while it’s thinking, and then jj diffedit to see the changes. All were simply hacks that were trying to fill what I was running from, learning. For some people it’s opening up twitter when their agent is working hence there more dev rants on the internet than they were before. But the thing is, as long as you are not writing code, the sadness will come up, impostor syndrome when you are congratulated for building a feature. Sure, it might work, but did you actually do it?

NEXT STEPS

I am planning to cleanse myself of this lack of touch with code by leaving agents and going back to the tab. The tab was somewhat perfect. I was good enough for you to know what you are doing and also good enough for it to write the code for you. It was the perfect abstraction.