Implementation:Truera Trulens TruGraph Wrapper

Knowledge Sources	TruLens TruLens LangGraph Guide
Domains	Observability, Agent_Evaluation
Last Updated	2026-02-14 08:00 GMT

Overview

Concrete tool for wrapping and instrumenting compiled LangGraph agents for agent-specific tracing and evaluation, provided by the trulens-apps-langgraph package.

Description

The TruGraph class is a specialized recorder for LangGraph applications. It inherits from TruChain and extends it with LangGraph-specific instrumentation for graph nodes, task functions, and Pregel execution methods. The wrapper automatically discovers and instruments LangGraph internal components.

TruGraph combines both LangChain and LangGraph instrumentation via a CombinedInstrument that merges module, class, and method discovery from both frameworks.

Usage

Import and use TruGraph when evaluating compiled LangGraph agents. Wrap the compiled graph (the result of graph.compile()), not the StateGraph builder. Configure with agent evaluation metrics (tool selection, relevance, groundedness).

Code Reference

Source Location

Repository: trulens
File: src/apps/langgraph/trulens/apps/langgraph/tru_graph.py
Lines: L210-1814 (class L210, __init__ L1586-1658)

Signature

class TruGraph(TruChain):
    def __init__(
        self,
        app: Any,
        main_method: Optional[Callable] = None,
        **kwargs: Any,
    ):
        """
        Args:
            app: A LangGraph application (Pregel, StateGraph, or custom class).
            main_method: Optional entry point callable.
            **kwargs: Additional arguments including:
                app_name (str): Application name.
                app_version (str): Version tag.
                feedbacks (List[Feedback]): Agent evaluation metrics.
                connector (DBConnector): Optional database connector.
        """

Import

from trulens.apps.langgraph import TruGraph

I/O Contract

Inputs

Name	Type	Required	Description
app	Any	Yes	Compiled LangGraph (Pregel/CompiledStateGraph)
main_method	Callable	No	Optional entry point
app_name	str	No	Application name (via kwargs)
app_version	str	No	Version tag (via kwargs)
feedbacks	List[Feedback]	No	Agent evaluation metrics (via kwargs)

Outputs

Name	Type	Description
return	TruGraph	Instrumented wrapper with graph-aware tracing

Usage Examples

Wrap Compiled LangGraph Agent

from trulens.apps.langgraph import TruGraph
from trulens.core.session import TruSession
from langgraph.graph import StateGraph

session = TruSession()

# Build and compile agent graph
graph = StateGraph(AgentState)
graph.add_node("agent", agent_node)
graph.add_node("tools", tool_node)
graph.add_conditional_edges("agent", should_continue, {"continue": "tools", "end": "__end__"})
graph.add_edge("tools", "agent")
graph.set_entry_point("agent")
compiled_graph = graph.compile()

# Wrap compiled graph
tru_agent = TruGraph(
    compiled_graph,
    app_name="ReAct_Agent",
    app_version="v1",
    feedbacks=[f_tool_selection, f_relevance, f_groundedness],
)

# Record agent execution
with tru_agent as recording:
    result = compiled_graph.invoke({"messages": [("user", "What is TruLens?")]})

Related Pages

Implements Principle

Principle:Truera_Trulens_LangGraph_Agent_Wrapping

Requires Environment

Uses Heuristic

Heuristic:Truera_Trulens_Trace_Compression_Token_Limits

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment