LongSuLoongSuiteite LangGraph Instrumentation

LoongSuite instrumentation for LangGraph.

Installation

# Step 1: install LoongSuite distro
pip install loongsuite-distro

# Step 2 (Option C): install this instrumentation from PyPI
pip install loongsuite-instrumentation-langgraph

Usage

from opentelemetry.instrumentation.langgraph import LangGraphInstrumentor

LangGraphInstrumentor().instrument()

What it does

This instrumentation patches two targets to enable LangChain instrumentation (loongsuite-instrumentation-langchain) to recognise LangGraph ReAct agents and create proper Agent / ReAct Step spans.

All patches use wrapt.wrap_function_wrapper (consistent with loongsuite-instrumentation-langchain).

1. create_react_agent patch

Wraps langgraph.prebuilt.create_react_agent to set a boolean flag _loongsuite_react_agent = True on the compiled CompiledStateGraph. The flag itself is not accessible inside LangChain's callback system — it only serves as the trigger for the second patch below.

2. Pregel.stream / Pregel.astream patch

Wraps the graph execution entry points so that when a graph carrying _loongsuite_react_agent = True is invoked, the metadata {"_loongsuite_react_agent": True} is injected into the RunnableConfig before execution begins.

The data flow:

graph._loongsuite_react_agent = True       # set by patch 1
            │
            ▼
Pregel.stream() wrapper intercepts call    # patch 2
            │
            ▼
config["metadata"]["_loongsuite_react_agent"] = True
            │
            ▼
LangChain callback manager reads metadata
            │
            ▼
Run.metadata["_loongsuite_react_agent"]    # LoongsuiteTracer reads this

LangChain's callback system automatically propagates config["metadata"] to all child callbacks, so every sub-node within the graph also carries the flag. The LoongsuiteTracer disambiguates the top-level graph (Agent span) from child nodes (chain spans) by tracking an internal inside_langgraph_react flag that propagates through the run hierarchy.

Pregel.invoke() / ainvoke() internally delegate to stream / astream, so only the latter two need to be patched.

How it works with LangChain instrumentation

When both instrumentors are active, the LoongsuiteTracer in the LangChain instrumentation:

1. Detects the agent — _has_langgraph_react_metadata(run) checks Run.metadata for the flag. If the parent is not already inside a LangGraph agent, this run becomes an Agent span.

2. Resolves agent name — when the ReAct agent is invoked inside an outer graph node (e.g. product_agent), the agent span inherits the node's name (invoke_agent product_agent) instead of the generic default (invoke_agent LangGraph).

3. Tracks ReAct steps — each time the "agent" node fires inside the graph, a new ReAct Step span is created, with the hierarchy: Agent > ReAct Step > LLM / Tool.

Compatibility

• langgraph >= 0.2
• langchain_core >= 0.1.0
• Python 3.9+