Not sure I understood well your comment.

Do you propose just ask AI to generate orchestration in Python?

If it generates orchestration in python, then you can review it and improve it, and it's deterministic and efficient and doesn't cost tokens to nondeterministically and unreliably interpret every time. And you can even use any of the zillions of standard well supported python modules. Does your language have a rich library of modules like PyPI or CPAN or NPM? Or are you hoping to bootstrap that from a ghost town?

The overwhelming financially and environmentally costly problem of this approach of inventing new languages for LLMs to use, which sweeps all other lower order problems aside, is that you have to include the entire language definition in every prompt, which is paltry compared to the entire body of manuals, tutorials, source code, example code, books, courseware, stackoverflow questions, hn discussions, etc, about Python or any other popular language, that is already built into the model before the first token you send it. It's not worth arguing about any other issues, because that one is so overwhelmingly fatal.

Besides that, Mrs. Lincoln, how did you enjoy the play?