The way I see it the only jobs you are realistically automating (without full on AGI) are the CRUD and pipework jobs which are unambiguously explainable without a lot of context.
The difficulty/complexity in delivering software is mostly on having accurate domain knowledge on the problem you are trying to solve and how users or other machines in that environment interact with said software.
So again, unless we hit a landmark of a very outperforming AGI I think that much of the domain knowledge in many software 'niches' will be out of reach of automation.
That is not to say that many sub-tasks/boilerplate cannot be automated.
But you are missing the point that even a robot could code 90% of the boilerplate and leave it to someone with domain knowledge to check it through and possibly tweak a few bits.
The difficulty/complexity in delivering software is mostly on having accurate domain knowledge on the problem you are trying to solve and how users or other machines in that environment interact with said software.
So again, unless we hit a landmark of a very outperforming AGI I think that much of the domain knowledge in many software 'niches' will be out of reach of automation.
That is not to say that many sub-tasks/boilerplate cannot be automated.