For example a lot of people would find writing scripts cumbersome, but not a person who's written a lot of them. They're not any more fragile than other logic error capable software is
With k8s, a horizontal autoscaler is a few lines in a yaml file and the result works in any cluster run by any vendor.
No tool "works the rest out". It's _always_ a compromise, because inherent complexity can never be removed, only moved. What you gain in one area, encumbers another.
You may freely use k8s, but it's not magical nor easier to use than any existing systems. In fact adopting it often takes non-trivial time and the web is full of failure stories with very benign warnings and catastropic results.
Maybe I'm spoiled by GCP and it's not the same in other providers; but I can have a brand new debian VM configured and operational in less than 10s (less the time that I run my own startup script).
Debian machines spawn with incredible speed; not much slower (if at all slower) than a new container.
There are also many options for creating Debian based image files with all you need baked in.
I guess it depends very much on the application you have.
I think I've been in total of one company where it was somewhat true, and even there we didn't do it.