what is a good way to test agentic workflow/pipeline? i'm looking for caching or mocking so that the whole agentic workflow can be tested without llm calls? something similar to https://docs.langchain.com/oss/python/langgraph/graph-api#node-caching

What is the main part of the system you want to test, the LLM or the harness? BAML test blocks are great for the former. For the latter, I’d consider doing end to end integration tests with the real LLM. It’s likely not as expensive as your prod traffic, and it keeps the real performance visible to you by exposing your test suite to the same indeterminism that prod sees. If you are mostly trying to test your harness, record-relay or other http mocking solutions are nice and deterministic. I don’t think they need to be LLM-specific. For one of my projects I mock out the LLM provider as if it were any other REST endpoint. (If you need streaming, the story is more complicated)