LangChain approach struck me as interesting, but I never really saw much inherent utility in it. For our production code we went with direct use of LLM runtime libraries and it was more than enough.
We've had success with non-developers using some of the visual tools built on top of LangChain to build exploratory models in order to prove a concept. LangChain does seem well suited to providing the "backend" for that type of visual node-based modelling.
Of course, once the model is proven it is handed off to developers to build something more production-worthy.
