With LLMs you say “you want to do X” and voila, personality.

What is indeed missing from current implementations is continuous looping. Doing actions and taking stock of the results. I guess that’s kind of expensive right now. We’ll get there. I don’t see the fundamental problem.