The answer to the first question is it depends. I've found that a well tuned assistant can handle branching and extraction as well as if not better than a hastily written workflow. The big advantage I see with workflows that I still need to fully test is the ability to switch models and other parameters while building a multilingual assistant. This advantage is what got me playing with workflows but unfortunately I haven't had the time to really dig into and test it. If it works then I can see that fixing my last few remaining pain points with model latency due to compromising on choosing models optimised for multilingual versus exploiting the correct model based on the current language being used