Matthew Stafford

Update on previous LangChain post: LangChain bad. Swapped out for a simpler wrapper library (litellm) and the transitions library in lieu of LangGraph to handle the state machine. Way, way, way better.

There are whole classes built around things that are much more easily done, and way more understandable, with basic f-string interpolation of prompts.

I've been working with LangChain for a little bit now and I really do understand the hate. It's one too many layers of abstraction for no real gain. LangGraph is the best part of it but you could implement the same thing with a much thinner wrapper around the LLM APIs.

This isn't to shit on the space - more to say that doing data and tech work in a mixed small-p/big-P political environment with limited budgets is *hard*. Very little of all that money you see donated is going to the tech.

Convinced that if more people saw how the sausage was made behind the tech that drives political outreach/persuasion/GOTV you'd see that there are no conspiracies; we're still working on how to collaborate sanely.

You all realize agents are just orchestrated microservices right?