I definitely understand where you are coming from. I don't understand why not so many people consider this angle. If the coding part of your job is to be replaced by being a code reviewer, then you should be good at code review and you should understand the code that is being produced on the same deep level, which is expertise that only comes with time.

But how do you maintain that deep understanding of code and be better at code review? You didn't learn how to code by just reading code, you wrote code to understand what works and what doesn't work. If you don't want to sloppify your codebase you need to perform quality code review, but to perform quality code review you need to ensure that your skills stay sharp through writing code. So a hybrid approach of using AI starts to make more sense, where you automate tasks which are slam dunks for AI agents, and continue doing deep work on stuff that is more interesting and fun to you.

I know to some companies it really doesn't matter how fun your work is to you, the only metric that matters is productivity, but that productivity will fall in quailty if we always offload and don't maintain our understanding of code.

What happens if Claude hikes up the token prices after they stop being subsidized, these models are only getting more expensive so far? Companies will be forced to submit.

Single point of failure, we all become reliant on a select few companies to provide our skills back to us that we spent a lot of time learning, imagine if they just went offline and at the same time you had to fix prod? There is still importance in knowing how to code.

My agentic workflow involves delegating tasks to Claude Code either on another worktree, or the same worktree, sometimes a hybrid of working on something manually then filling in the gaps somewhere else, or being fully manual / AI as better google search on one task.