I am aware of two approaches to building something like this: Either through a REPL-like prompt chaining approach (like langchain), or a toolformer-like approach where tool use becomes a trainable choice. Any sources/speculations on which it could be (or something else)? What do you all think? Thanks.