I'm not convinced there is a missing part. It's hard to summarize but to use a metaphor, it's like wanting a self-driving car that crashes sometimes to go faster and faster. The missing piece is just the AI figuring out what it needs to figure out. In theory we're heading towards every single tool becoming obsolete as AI can understand and do more... but that's in theory of course. If scaling LLM transformers does not deliver, then yes perhaps some better tooling will help, but I still think it has some fundamental flaws that in the end are not saving that much time when all the mistakes and corrections are factored in.