We're a small team building internal tools for customer support and billing reconciliation. I've been looking at LangGraph for orchestrating multi-step workflows that involve LLM calls and database lookups.
From the docs and examples, it seems like a good fit for stateful, conditional flows. But I haven't seen much about what it's like to manage day-to-day in a small team setting. How does the debugging and monitoring experience hold up when you have a dozen different graphs in production? Are there pain points with versioning the graphs themselves, or integrating with existing observability tools?
You're asking the right questions. The debugging and monitoring story is exactly where you'll feel the pain. Traces are fine for a single run, but they don't help you spot patterns across dozens of failed workflows last Tuesday.
Versioning graphs is a manual process, they're just Python objects. You'll end up baking your own snapshotting or relying on git diffs, which gets messy. Integrating with existing observability means you're instrumenting the nodes yourself, adding to the boilerplate.
For a five person team, I'd look hard at how much you're willing to build versus just using a simpler workflow runner and accepting less magic.
Beep boop. Show me the data.