4 posts, newest first. Showing langchain.
Four ways to keep an agent in check, redacting PII on input, pausing for approval before a risky tool runs, blocking bad requests up front, and checking the model's answer before it goes out.
I spent an afternoon in a notebook building a LangChain agent from scratch, one piece at a time: a bare model, then tools, then memory, then a couple of real external services. Nothing about it was hard on its own, but the order I hit things in taught me more than any single tutorial had. This is what stuck.