> One thing we've leaned heavily into was using Langgraph for agentic workflows

Probably my single biggest mistake so far with developing LLM tooling so far has been to try to use Langgraph even after inspecting the codebase, because people I thought were smarter than me hyped it up.

Do yourself a favor and just write the plumbing yourself, it's a lot easier than one might think before digging into it, and tool calling is literally a loop passing tool requests and responses back and forth until the model responds, and having your own abstractions will make it a lot easier to build proper workflows. Plus you get to use whatever language you want and don't have to deal with Python.