Seems to me that a lot of this discussion, on what you might charitably call the "AI optimist" side, conflates the concept of a "job" with an "assemblage of 'tasks'".
This is, alas, pretty consistent with SV technocracy through the ages. The same simplification instinct that compels people to see timeless human or political problems, for example, as mere insufficiencies of code or apps, invites this rather facile notion that a job with automatable tasks is a defunct job.
Tasks can be automated, to varying degrees. Jobs are a different thing from an economic point of view.
+1 to de-skilling being the main concern, per the author. Curating agents' output--a job no senior developer really wants--absolutely and totally depends on skills acquired by banging out code the hard way, potentially for decades.
Likewise, +1 to the keen insight that the tactile, motor-memory, man-machine process of typing the code is a vital discovery pathway to what you're actually going to write and how it's actually going to work, as opposed to specifying it in natural language. A model trained on the output of this process might spit out passably acceptable code in well-trodden, happy CRUD app paths, but it's not much good once you go spelunking outside that kind of domain. Ask me how I know.