How to use return_direct=True correctly in tools and control message appending behavior in LangGraph?

Hi everyone,

I’m trying to better understand how to use the return_direct=True option in LangGraph tools, especially when using them inside a ToolNode.

What I’m doing

  • I have several tools defined with @tool(return_direct=True).

  • These tools are used inside a ToolNode in my agent graph.

  • However, when the graph executes, I still see AI messages being generated after those tools run, even though I expected the tool output to be returned directly and stop message generation.

The issue

Some of my tools return very large results (e.g., get_all_products with full product details), and I don’t want those outputs:

  • To be added to the state messages, or

  • To trigger an extra AI message that just repeats the same tool output.

On the other hand, I have some tools where I do want the results appended to the state (for context), but still don’t want an AI message generated afterward.

What I know

I understand I could create a router node to detect specific tool names and handle this behavior manually (deciding when to return directly or not).
However, I’d prefer to keep using ToolNode for its built-in advantages and simplicity.

My questions

  1. How can I make return_direct=True actually prevent AI message generation after tool execution when using ToolNode?

  2. Is there a way to handle mixed behavior — for example:

    • Some tools return directly (display result only, no state append),

    • Some tools append their output to the state but skip the AI message?

Any examples or best practices for handling this pattern would be greatly appreciated!


Hi @rasoul111000

I’m not sure if this is what you’re looking for, but here’s my attempt:


================================ Human Message =================================

Add 3 and 4.
================================== Ai Message ==================================
Tool Calls:
add (2a3d7d85-7a0b-4f88-9e55-83dafba55a9b)
Call ID: 2a3d7d85-7a0b-4f88-9e55-83dafba55a9b
Args:
a: 3
b: 4
================================= Tool Message =================================
Name: add

7

I modified the quickstart example by adding an edge from tool_node to END.

# Step 1: Define tools and model

from langchain.tools import tool
from langchain.chat_models import init_chat_model


model = init_chat_model(
    "ollama:qwen3",
    temperature=0
)


from langgraph.prebuilt.tool_node import ToolNode
from langchain_core.tools import tool

@tool
def add(a: int, b: int) -> int:
    """Add two numbers."""
    return a + b

tool_node = ToolNode([add])



# Augment the LLM with tools
model_with_tools = model.bind_tools([add])
# Step 2: Define state

from langchain.messages import AnyMessage
from typing_extensions import TypedDict, Annotated
import operator


class MessagesState(TypedDict):
    messages: Annotated[list[AnyMessage], operator.add]
    llm_calls: int

# Step 3: Define model node
from langchain.messages import SystemMessage


def llm_call(state: dict):
    """LLM decides whether to call a tool or not"""

    return {
        "messages": [
            model_with_tools.invoke(
                [
                    SystemMessage(
                        content="You are a helpful assistant tasked with performing arithmetic on a set of inputs."
                    )
                ]
                + state["messages"]
            )
        ],
        "llm_calls": state.get('llm_calls', 0) + 1
    }


# Step 4: Define tool node

from langchain.messages import ToolMessage




# Step 5: Define logic to determine whether to end

from typing import Literal
from langgraph.graph import StateGraph, START, END


# Conditional edge function to route to the tool node or end based upon whether the LLM made a tool call
def should_continue(state: MessagesState) -> Literal["tool_node", END]:
    """Decide if we should continue the loop or stop based upon whether the LLM made a tool call"""

    messages = state["messages"]
    last_message = messages[-1]

    # If the LLM makes a tool call, then perform an action
    if last_message.tool_calls:
        return "tool_node"

    # Otherwise, we stop (reply to the user)
    return END

# Step 6: Build agent

# Build workflow
agent_builder = StateGraph(MessagesState)

# Add nodes
agent_builder.add_node("llm_call", llm_call)
agent_builder.add_node("tool_node", tool_node)


# Add edges to connect nodes
agent_builder.add_edge(START, "llm_call")
agent_builder.add_conditional_edges(
    "llm_call",
    should_continue,
    ["tool_node", END]
)

agent_builder.add_edge("tool_node", END)
# Compile the agent
agent = agent_builder.compile()



# Invoke
from langchain.messages import HumanMessage
messages = [HumanMessage(content="Add 3 and 4.")]
messages = agent.invoke({"messages": messages})
for m in messages["messages"]:
    m.pretty_print()

If this isn’t what you meant, could you please share your code so we can understand better?