At first I thought this was going to be some LLM thing. I had an idea a while ago, "context-based module development", that I started on a prototype of but never followed up on. The idea is to have a standard format for defining modules, black boxes with interfaces and clear definitions, that can be composed hierarchically. Module definitions and their code should not be larger than the context window of an LLM. As long as each module is well defined and tested and treated like a black box, you could have a system composed of both human and AI built modules that should behave as expected and be somewhat comprehensible. Not all architectures would work with this and I don't know if it would have worked in the end, but I do expect that at some point a more formal system for defining software from the ground up for AI development will emerge.