Missing Input/Output Traces in Langfuse with LangGraph Agent Integration

[Dodero10]

I'm currently using LangGraph to build and run agents, and integrating Langfuse for tracing and observability. However, I've encountered an issue where the input and output of the agent steps are not being properly displayed or captured in Langfuse during tracing. This makes it difficult to debug or analyze the behavior of the agents effectively.

I've verified that tracing is initialized and Langfuse is recording events, but the structured input/output for each step (e.g., tool calls, thoughts, actions) isn't showing up as expected.

Has anyone else experienced this when using LangGraph with Langfuse? Any suggestions or best practices for making sure input/output is properly tracked in this setup would be appreciated.

Input and Ouput show metadata

This is my code

async def main():
    
    test_query = "Cho tôi biết Thống kê thị trường trái phiếu doanh nghiệp riêng lẻ tháng 4/2025"
    
    print("==== TESTING REACT AGENT ====")

    async for s in react_graph.astream({
        "messages": [HumanMessage(content=test_query)]
    }, config={"callbacks": [langfuse_handler]}):
        print(s)
        print("--------------------------------")

from typing import Dict, List, Literal, cast

from langchain_core.messages import AIMessage
from langgraph.graph import StateGraph
from langgraph.prebuilt import ToolNode

from agents.configuration import Configuration
from agents.state import InputState, State
from agents.tools import TOOLS
from agents.utils import load_chat_model

async def react_agent(state: State) -> Dict[str, List[AIMessage]]:
    """Call the LLM powering our "agent".

    This function prepares the prompt, initializes the model, and processes the response.

    Args:
        state (State): The current state of the conversation.
        config (RunnableConfig): Configuration for the model run.

    Returns:
        dict: A dictionary containing the model's response message.
    """
    configuration = Configuration.from_context()

    model = load_chat_model(configuration.model).bind_tools(TOOLS)
    system_message = configuration.system_prompt.format(
        system_time=datetime.now(tz=UTC).isoformat()
    )

    response = cast(
        AIMessage,
        await model.ainvoke(
            [{"role": "system", "content": system_message}, *state.messages]
        ),
    )

    if state.is_last_step and response.tool_calls:
        return {
            "messages": [
                AIMessage(
                    id=response.id,
                    content="Sorry, I could not find an answer to your question in the specified number of steps.",
                )
            ]
        }

    return {"messages": [response]}


builder = StateGraph(State, input=InputState, config_schema=Configuration)

builder.add_node(react_agent)
builder.add_node("tools", ToolNode(TOOLS))

builder.add_edge("__start__", "react_agent")


def route_model_output(state: State) -> Literal["__end__", "tools"]:
    """Determine the next node based on the model's output.

    This function checks if the model's last message contains tool calls.

    Args:
        state (State): The current state of the conversation.

    Returns:
        str: The name of the next node to call ("__end__" or "tools").
    """
    last_message = state.messages[-1]
    if not isinstance(last_message, AIMessage):
        raise ValueError(
            f"Expected AIMessage in output edges, but got {type(last_message).__name__}"
        )
    # If there is no tool call, then we finish
    if not last_message.tool_calls:
        return "__end__"
    # Otherwise we execute the requested actions
    return "tools"

builder.add_edge("tools", "react_agent")
builder.add_conditional_edges(
    "react_agent",
    route_model_output,
)
# Compile the builder into an executable graph
graph = builder.compile(name="ReAct Agent")

[dosubot[bot]]

Hello @Dodero10! I'm Dosu, your friendly bot here to help you troubleshoot bugs, answer questions, and navigate your contributions. I'm ready to assist you while you wait for a human maintainer. How can I help you today?

To ensure that the input and output traces of agent steps are properly displayed and captured in Langfuse when using LangGraph, you should explicitly set the output using langfuse.trace() after your generation call. Here's an example of how you can modify your code to achieve this:

async def react_agent(state: State) -> Dict[str, List[AIMessage]]:
    configuration = Configuration.from_context()
    model = load_chat_model(configuration.model).bind_tools(TOOLS)
    system_message = configuration.system_prompt.format(
        system_time=datetime.now(tz=UTC).isoformat()
    )

    response = cast(
        AIMessage,
        await model.ainvoke(
            [{"role": "system", "content": system_message}, *state.messages]
        ),
    )

    # Capture the output trace
    langfuse.trace(id=state.trace_id, output=response.content)

    if state.is_last_step and response.tool_calls:
        return {
            "messages": [
                AIMessage(
                    id=response.id,
                    content="Sorry, I could not find an answer to your question in the specified number of steps.",
                )
            ]
        }

    return {"messages": [response]}

Make sure you are using Langfuse version 2.57.3 or later, as this version supports capturing input/output data effectively ^[1]. If you are using the Traceloop OpenLLMetry SDK and experiencing missing input/output data, consider using the Langfuse SDK directly to capture input/output data for LangGraph workflow spans ^[2].

To continue talking to Dosu, mention @dosu.

---

^{Help Dosu learn! Give it feedback: Great Response | Irrelevant Answer | Incorrect Sources | Too Verbose | Hallucination | Bug Report | Other}

[Leo310]

I'm facing a similar issue, where traces are tracked, but the inputs and outputs aren't being visualized properly.

This seems linked to #6330 and #7043. Specifically, Langgraph messages appear to use a "type" attribute instead of the "role" attribute required by the ChatML schema, which likely causes the rendering issue.

It would significantly improve usability if Langgraph messages could also be displayed in a user-friendly, formatted way, similar to other messages, rather than as raw json messages.

[bridger217]

+1, This is happening to me when using the OpenAI Agents SDK. Example trace: 4048f8198f58476fab84582e10c66567

[ktwillcode]

any solution yet?

[MaxiWeissenbacher]

+1

[gabadi]

+1

[markine]

+1 I've hacked something together by translating attributes in custom spans, but it's ugly. We know that Langfuse, Arize, Braintrust and others are legitimate processors of agent interaction logs, so Livekit should cleanly deliver data to them. Notwithstanding development of your own Agent Insights features. Thank you!