For the sake of the argument, please assume that we will arrive at a point where all code changes of more than a few characters can be performed more quickly by an AI.

By saying "more quickly", I would still allow for the fact that the AI may make mistakes. I would simply assume that the overall process up to committing the code, which may include fixing mistakes, is faster, on average, than writing things by hand.

In this scenario, there could still be various reasons why writing code by hand remains a part of software development. I can think of two:

1. There could be some kind of "embodied understanding" of code, some kind of "embodied cognition" [2]. We might imagine that there is some class of coding mistakes that can only be avoided with a high chance by writing the code oneself. There could also be some pathway of thinking that remains much harder to activate by just reviewing code. Or there could be a limit to how well most humans can load large amounts of code into their "working memory" without them touching said code and "feeling" its reaction. Such effects could make writing code by hand a necessity, even if partially just to remain a sharp reviewer of generated code.

2. There could also be an almost cynical, economical reason. Let's assume that a certain share of all existing developers can simply not adopt to a different mode of working, one in which they run agents instead of coding. Let's say that share is 30% of today's workforce. As these are likely to be some of the more experienced and efficient developers, the job market could remain dependent on them for many years, even if those who generate code see a significant productivity boost.

Maybe you can think of other reasons. In any case, I am very interested in your thoughts.

[1] https://news.ycombinator.com/item?id=47998225 [2] https://plato.stanford.edu/entries/embodied-cognition/