This is the right balance for me as well.

I use an agent to generate a first-pass attempt, and then (deadlines willing), I manually read every line at least once so I understand what the code actually does.

Then I manually fix the inevitable slop that is mixed in with the good stuff, and only once the code is up to my personal standards do I send it.

This probably reduces my “AI performance boost” to 30-50% instead of the huge gains reported by others. But I retain the ability to reason about the codebase and use AI much more precisely when I’m trying to troubleshoot production outages or subtle bugs — something I notice the rest of my team struggles with, since adopting “agentic workflows” everywhere.

I think actively working to retain some cognitive flexibility and “muscle memory” around coding tasks is going to be rather advantageous in the long run.