I think LangChain serves its purpose when you are new to AI and LLM and have to experiment a lot. That was my case. The landscape is changing so fast it is hard to keep up. With LangChain I was able to try different LLM and their features with minimal efforts. Granted, I haven't taken anything to production yet. But using LangChain to narrow down to the LLM you like, then you can use the LLM directly without going through LangChain.