Subgraphs as Tools

[dhirajsuvarna]

Is it possible to use subgraphs as tools? If yes, any guidance on how to do it will be helpful.

[bedead]

I am not sure if there is direct implementation of doing this, but you can create a simple workaround like this.

from email_agent.graph import graph


@tool
def email_graph_tool(input: str):
    """
    Tool calling for runing email_graph
    Parameters:
        input: str -> whatever this input is
    """
    config = {}
    result = graph.invoke(input={"input": input}, config=config)

    return result

and now you can add this tool to your main graph or create_react_agent.
this will likely work.
if not I am open to discuss.

[marekratho]

Hi, I'm trying to use this pattern for adding a graph as a tool for a node within another graph. I'm running into a problem that the graph I want to use as a tool contains an interrupt call and after submitting the resume command the execution restarts in the tool rather than in the node which contains the interrupt call. Do you have any experience with this/advice on how to solve it?

[bedead]

@marekratho Again I am not aware of any perfect solution for this but, one trick you can try here is that you can add a new state attribute in your subgraph which you what to run as tool, the attribute could be something like a Boolean value or enum type for better customization. And then at start of subgraph you can check whether this state attribute is changed or not if yes you can directly route to resume node and continue working from there, else run the default route.

I assume as you are using resume in your subgraph, it must have checkpoint which will store the data for a consistent runs for a while. This implementation will probably work even in execution starts from tool node.

[bedead]

@marekratho also if you can share some code snippet/ with debug logs, it would be more helpful to understand the problem and find the solution 🙂.

[atom2ueki]

I ran into same issues, willing to make llm automatically call subgraph as tool, but this way doesn't works, at the end I fallback to invoke suggraph manually by checking a empty tool calling

[dhirajsuvarna]

@bedead - this is exactly what I ended up doing.

Though there are two trivial issues that I would like to highlight -

1. MessageList is populated with ToolMessages instead of AIMessages
2. There isn't nice visualization (mermaid diagrams) for Graphs which are used as Tools inside the main graph

[bedead]

@dhirajsuvarna - Another way could be to create a custom tool node instead and in that we do basically the same thing, but we have to manage the execution part of the tool by ourselves.

something like this

# This will be list of subgraphs you have
tools_by_name = {"researchGraph": researchGraph, "writerAgent" :writerAgent}

def custom_tool_node(state: AgentState):  
    outputs = []  
    for tool_call in state["messages"][-1].tool_calls:  
        tool_result = tools_by_name[tool_call["name"]].invoke(tool_call["args"])  
        # Instead of ToolMessage, create AIMessage  
        outputs.append(  
            AIMessage(  
                content=f"Tool {tool_call['name']} returned: {json.dumps(tool_result)}",  
                # Optionally include metadata about the tool call  
            )  
        )  
    return {"messages": outputs}

This will fix "1. MessageList is populated with ToolMessages instead of AIMessages". But for fixing 2. you have to invoke each subgraph manually with conditional edges or custom logic for node execution.

[dhirajsuvarna]

I think this should be raised as a PR in langgraph library, so that it can be available to all, also I am reluctant to use custom ToolNode because, the one provided by library, gives lot of functionality inclusion error handling and retry mechanism.