I have experimented in a small way with having my APIs return requested data AND instructions back to my Custom GPT on how to process and/or display that data. And it works very well. But I’ve wondered whether the entire GPT could be run this way.
You’d have just very simple hard-coded instructions in the GPT, asking it to call an endpoint. The endpoint would return data and instructions. The GPT would follow those instructions and so on.
Has anyone done that? Sitting down to try myself right now.