Thanks for the question! I could maybe see that being the case for LangChain, specifically, if you’re just thinking about abstracting the LLM provider’s APIs.

Thanks for the question! I could maybe see that being the case for LangChain, specifically, if you’re just thinking about abstracting the LLM provider’s APIs. But here, we’re specifically using LangGraph to handle checkpointing and state management. This is a much more difficult challenge to get right in a sufficiently complex application like Pizza Bot. Case in point: for the internal-only version, we wrote this ourselves. It works for the most part, but there are subtle edge cases and bugs that we’ve had to tackle like a game of whack-a-mole. For the open source release, we looked at a few frameworks and deliberately settled on LangGraph (and DeepAgents). I’ve never been so happy to throw away code. :) In our experience so far, it’s been rock solid. Most of the issues/bugs we’ve had have come from streaming to the frontend, so a simple Cmd-R / Ctrl-R will get you back up and running without having to worry about your data being lost.