Appearance

LangChain ​

examples/05_langchain.py demonstrates the LangChain integration. The SDK's callback handler registers itself through LangChain's register_configure_hook(..., inheritable=True), so it is attached to every run without anyone passing callbacks=[...] — and a tool nested inside a chain inherits it from the parent run rather than needing its own wiring.

bash
python examples/05_langchain.py
AGENTSIGHT_EXAMPLES_OFFLINE=1 python examples/05_langchain.py

Offline it uses LangChain's own fake chat model: the callback path, the run tree and every span are real. With OPENAI_API_KEY set and langchain-openai installed it uses a real model and installs the OpenAI patch, which shows the stand-down rule — the handler emits no llm span for a provider a patch already covers, so each call is counted once, not twice.

What it runs ​

  • A turn with two tool calls and a chain — @tool-decorated LangChain tools invoked directly, then a RunnableLambda | model chain, so the spans nest the way a real application's do: the work inside the chain's run inherits the handler from the parent.
  • A failing tool — the span records the error.
  • Turnless spend — a chain invoked outside any turn: spend that belongs to the conversation but to no single exchange.

What to look for in the output ​

In examples/traces/05_langchain/: tool spans produced by LangChain's own @tool decorator with no AgentSight decorator in sight, and the chain's LLM call attributed correctly — once — whether the model is fake or real.