The Microsoft Research paper assessing intelligence capability of GPT4 effectively did this. If you just define APIs for the model to use under certain conditions it will write the API call. Once you do that, it's straightforward for a layer on top to detect the API call, actually execute it, and write the result back.