I don't think LLMs are going to change it as much as the massive concurrency that modern hardware provides. A single threaded task can only reach a tiny fraction of the CPU resources in a machine these days.

Things in the generated code (function names, variables) will be named better. Comments will be written with specific instructions about how to generate the following code. Test coverage will improve, but even fewer people will evaluate whether the tests are doing anything useful.

I think structure will shrink and simplify to fit context window token limits at the expense of developer(human) experience and developer bias.