isn't this happening already? There's the concept called "thinking" where the models talks with itself before giving you the final answer