I built my first commercial LLM agent back in October/November last year. As a newcomer to the LLM space, every tutorial and youtube video was about using LangChain. But something about the project had that "bad code" smell about it.
I was fortunate in that the person I was building the project for was able to introduce me to a few other people more experienced with the entire nascent LLM agent field and both of them strongly steered me away from LangChain.
Avoiding going down that minefield ridden path really helped me out early on, and instead I focused more on learning how to build agents "from scratch" more or less. That gave me a much better handle on how to interact with agents and has led me more into learning how to run the various models independently of the API providers and get more productive results.
I've only ever played around with it and not built out an app like you have, but in my experience the second you want to go off script from what the tutorials suggest, it becomes an impossible nightmare of reading source code trying to get a basic thing to work. LangChain is _the_ definition of death by abstraction.
I have read the whole source of LangChain in Rust (there are no docs anyway), and it definitely seems over-engineering. The central premise of the project, of complicated chains of prompts is not useful to many people, and not to me either.
On the other hand it took some years into the web, for some web frameworks to emerge and make sense, like Ruby on Rails. Maybe in 3-4 years time, complicated chains of commands to different A.I. engines will be so difficult to get right that a framework might make sense, and establish a set of conventions.
Agents, another central feature of LangChain, are not proved to be very useful as well, for the moment.
LangChain got its start before LLMs had robust conversational abilities and before the LLM providers had developer decent native APIs (heck, there was basically only OpenAI at that time). It was a bit DOA as a result. Even by last spring, I felt more comfortable just working with the OpenAI API than trying to learn LangChain’s particular way of doing things.
Kudos to the LangChain folks for building what they built. They deserve some recognition for that. But, yes, I don’t think it’s been particularly helpful for quite some time.
I tried to use Langchain a couple times, but every time I did, I kept feeling like there was an incredible amount of abstraction and paradigms that were completely unnecessary for what I was doing.
I ended up calling the model myself and extracting things using a flexible json parser, I ended up doing what I needed with about 80 lines of code.
This is their game. Infiltrate HN, X, YouTube, Google with “tutorials” and “case studies”. Basically re-target engineers until they’ve seen your name again and again. Then, they sell.
I was fortunate in that the person I was building the project for was able to introduce me to a few other people more experienced with the entire nascent LLM agent field and both of them strongly steered me away from LangChain.
Avoiding going down that minefield ridden path really helped me out early on, and instead I focused more on learning how to build agents "from scratch" more or less. That gave me a much better handle on how to interact with agents and has led me more into learning how to run the various models independently of the API providers and get more productive results.