Inconsistent history when using tool that returns Command with goto=END and graph=Command.Parent

OSS Product Help LangGraph
1

Hey,

I’m currently facing an issue with a supervisor agent that handoffs work to another agent which ultimately calls a tool. I would like this tool to have a set output structure, since the output is fairly large and structured, if I return the tool output back to the agent, it always modifies it in some way.

Here is a simplified example of my problem:

from langchain.agents import AgentState, create_agent
from langchain.chat_models import init_chat_model
from langchain_core.messages import HumanMessage, ToolMessage
from langchain_core.runnables import RunnableConfig
from langchain_core.tools import tool
from langgraph.checkpoint.memory import InMemorySaver
from langgraph.graph import END, START, StateGraph
from langgraph.prebuilt import ToolRuntime
from langgraph.types import Command

model = init_chat_model("openai:gpt-4.1")


@tool
def weather_tool(city: str) -> str:
    """
    Returns what's the weather in the specified city
    """
    return f"It's sunny in {city}"


@tool
def end_weather_tool(runtime: ToolRuntime, city: str) -> Command:
    """
    Returns what's the weather in the specified city
    """
    return Command(
        graph=Command.PARENT,
        goto=END,
        update={
            "messages": [
                ToolMessage(f"It's sunny in {city}", tool_call_id=runtime.tool_call_id)
            ]
        },
    )


def create_handoff_tool(*, agent_name: str, description: str | None = None):
    name = f"transfer_to_{agent_name}"
    description = description or f"Ask {agent_name} for help."

    @tool(name, description=description)
    def handoff_tool(runtime: ToolRuntime) -> Command:
        tool_message = {
            "role": "tool",
            "content": f"Successfully transferred to {agent_name}",
            "name": name,
            "tool_call_id": runtime.tool_call_id,
        }
        return Command(
            goto=agent_name,
            update={
                "messages": runtime.state["messages"] + [tool_message],
            },
            graph=Command.PARENT,
        )

    return handoff_tool


weather_agent = create_agent(
    model=model,
    system_prompt="You are a weather agent",
    tools=[
        # end_weather_tool,
        weather_tool,
    ],
    name="weather_agent",
)

supervisor_agent = create_agent(
    model=model,
    system_prompt="You are a supervisor, handoff weather questions to the weather_agent",
    tools=[
        create_handoff_tool(agent_name="weather_agent"),
    ],
    name="supervisor_agent",
)


graph = (
    StateGraph(state_schema=AgentState)
    .add_node(supervisor_agent, destinations=(END, "weather_agent"))
    .add_node(weather_agent, destinations=(END,))
    .add_edge(START, "supervisor_agent")
)

compiled_graph = graph.compile(
    checkpointer=InMemorySaver(),
    name="graph",
)

out = compiled_graph.invoke(
    {"messages": [HumanMessage(content="What's the weather in Sydney?")]},
    config=RunnableConfig({"configurable": {"thread_id": "thread1"}}),
)

for msg in out["messages"]:
    msg.pretty_print()

If the agent calls the weather_tool, the history looks correct but the message gets rewritten by the agent → supervisor does the handoff, weather_agent calls the tool and the tool call is recorded in the history

WITH weather_tool
================================ Human Message =================================

What’s the weather in Sydney?
================================== Ai Message ==================================
Tool Calls:
transfer_to_weather_agent (call_H67T0KRvBs8q6EB2ZjFWyEBd)
Call ID: call_H67T0KRvBs8q6EB2ZjFWyEBd
Args:
================================= Tool Message =================================
Name: transfer_to_weather_agent

Successfully transferred to weather_agent
================================== Ai Message ==================================
Tool Calls:
weather_tool (call_a3KFnu3lAG0YU4I85rmJ7iZL)
Call ID: call_a3KFnu3lAG0YU4I85rmJ7iZL
Args:
city: Sydney
================================= Tool Message =================================
Name: weather_tool

It’s sunny in Sydney
================================== Ai Message ==================================

It is currently sunny in Sydney. If you need more details or a forecast, let me know!

If the agent calls the end_weather_tool, that uses Command with goto=END and graph=Command.PARENT the history doesn’t have the AIMessage that contains the tool call

WITH end_weather_tool
================================ Human Message =================================

What’s the weather in Sydney?
================================== Ai Message ==================================
Tool Calls:
transfer_to_weather_agent (call_PS0Vphj20wWp0RooZYWu8umC)
Call ID: call_PS0Vphj20wWp0RooZYWu8umC
Args:
================================= Tool Message =================================
Name: transfer_to_weather_agent

Successfully transferred to weather_agent
================================= Tool Message =================================
Name: end_weather_tool

It’s sunny in Sydney

The next call with this history will fail with following error:

openai.BadRequestError: Error code: 400 - {‘error’: {‘message’: “Invalid parameter: messages with role ‘tool’ must be a response to a preceeding message with ‘tool_calls’.”, ‘type’: ‘invalid_request_error’, ‘param’: ‘messages.[12].role’, ‘code’: None}}

Am I doing something that I’m not supposed to?
Is it fine to return AIMessage from the tool instead of ToolMessage as this doesn’t corrupt the history?

Would appreciate if someone proposed a better approach if there is any, thank you :slight_smile: !

EDIT: I’m not using return_direct=True because when I do the exceptions are not handled by my@wrap_tool_call middleware which is omitted from the example.

2

in:

@tool
def end_weather_tool(runtime: ToolRuntime, city: str) -> Command:
    """
    Returns what's the weather in the specified city
    """
    return Command(
        graph=Command.PARENT,
        goto=END,
        update={
            "messages": [
                ToolMessage(f"It's sunny in {city}", tool_call_id=runtime.tool_call_id)
            ]
        },
    )

why didn’t you update the message like this:

"messages": runtime.state["messages"]
            + [
                ToolMessage(f"It's sunny in {city}", tool_call_id=runtime.tool_call_id)
            ]
3

hi @hsm207

it’s because of the fact that create_agent produces an agent with its state that has the reducer add_messages applied to messages. messages: Required[Annotated[list[AnyMessage], add_messages]].

4

It seems you’ve stumbled on an edge case (or maybe this is by design?).

The AIMessage with tool_calls gets dropped when the tool returned just a ToolMessage in the Command, which OpenAI’s API needs as context.

The interim solution would be to explicitly include the AIMessage that triggered the tool call, along with your ToolMessage, in the update dictionary.

5

If the agent calls the weather_tool , the history looks correct but the message gets rewritten by the agent

I kinda think this is a feature of create_agent now. It’s meant for standard agentic loops where you want the agent to be helpful and conversational.

So, to prevent the Agent from summarizing the tool call would go against the intended behavior.

Therefore, it’s better to build the graph yourself to enforce your specific requirements for state management.

For example, see here.

6

Firstly, as suggested by @hsm207, the following code solves the problem:

from langchain.agents import AgentState, create_agent
from langchain.chat_models import init_chat_model
from langchain_core.messages import HumanMessage, ToolMessage
from langchain_core.runnables import RunnableConfig
from langchain_core.tools import tool
from langgraph.checkpoint.memory import InMemorySaver
from langgraph.graph import END, START, StateGraph
from langgraph.prebuilt import ToolRuntime
from langgraph.types import Command
from dotenv import load_dotenv

load_dotenv()

model = init_chat_model("openai:gpt-5-nano")


@tool
def weather_tool(city: str) -> str:
    """
    Returns what's the weather in the specified city
    """
    return f"It's sunny in {city}"


@tool
def end_weather_tool(runtime: ToolRuntime, city: str) -> Command:
    """
    Returns what's the weather in the specified city
    """
    return Command(
        graph=Command.PARENT,
        goto=END,
        update={
            "messages": runtime.state["messages"] # modification is here
            + [ToolMessage(f"It's sunny in {city}", tool_call_id=runtime.tool_call_id)]
        },
    )


def create_handoff_tool(*, agent_name: str, description: str | None = None):
    name = f"transfer_to_{agent_name}"
    description = description or f"Ask {agent_name} for help."

    @tool(name, description=description)
    def handoff_tool(runtime: ToolRuntime) -> Command:
        tool_message = {
            "role": "tool",
            "content": f"Successfully transferred to {agent_name}",
            "name": name,
            "tool_call_id": runtime.tool_call_id,
        }
        return Command(
            goto=agent_name,
            update={
                "messages": runtime.state["messages"] + [tool_message],
            },
            graph=Command.PARENT,
        )

    return handoff_tool


weather_agent = create_agent(
    model=model,
    system_prompt="You are a weather agent",
    tools=[
        end_weather_tool,
        # weather_tool,
    ],
    name="weather_agent",
)

supervisor_agent = create_agent(
    model=model,
    system_prompt="You are a supervisor, handoff weather questions to the weather_agent",
    tools=[
        create_handoff_tool(agent_name="weather_agent"),
    ],
    name="supervisor_agent",
)


graph = (
    StateGraph(state_schema=AgentState)
    .add_node(supervisor_agent, destinations=(END, "weather_agent"))
    .add_node(weather_agent, destinations=(END,))
    .add_edge(START, "supervisor_agent")
)

compiled_graph = graph.compile(
    checkpointer=InMemorySaver(),
    name="graph",
)

out = compiled_graph.invoke(
    {"messages": [HumanMessage(content="What's the weather in Sydney?")]},
    config=RunnableConfig({"configurable": {"thread_id": "thread1"}}),
)

for msg in out["messages"]:
    msg.pretty_print()

. Now, that begs the question: why do you need to pass the entire history if messages already has an add reducer, as @pawel-twardziak pointed out?

Well, when you use the Command to return execution to the Parent graph, the supervisor_agent has no visibility of the weather_agent’s messages. Therefore, all it sees is the tool message end_weather_tool after it called transfer_to_weather_agent.

For this reason, you need to manually return the chosen list of messages to the Parent graph. This is also why the implementation of the handoff_tool also returns the entire message history.

1 Like
7

If the agent calls the weather_tool , the history looks correct but the message gets rewritten by the agent

I kinda think this is a feature of create_agent now. It’s meant for standard agentic loops where you want the agent to be helpful and conversational.

So, to prevent the Agent from summarizing the tool call would go against the intended behavior.

Therefore, it’s better to build the graph yourself to enforce your specific requirements for state management.

For example:

Technically, you don’t actually need to return the entire history. You can just return the AIMessage that says there’s a tool call and the ToolMessage.

2 Likes