setting add_handoff_messages=False creates looping

[ashburncrash]

Hi team,

I'm using langgraph-supervisor 0.0.20 version. When passing add_handoff_messages=False and test it on LangGraph Studio, I see that there's a loop being created between supervisor node and subagent node. This loop doesn't occur if I remove add_handoff_messages=False.

orchestrator = create_supervisor(
        [subgraph1, subgraph2],
        model=model,
        state_schema=BaseAgentState,
        prompt=AGENT_PROMPT,
        supervisor_name="orchestrator",
        output_mode="last_message",
        add_handoff_messages=False,
    )

Is this expected behaviour? (not sure if you can see the run, but sharing the link to one of the runs that showed this: https://eu.smith.langchain.com/public/a4950620-1d85-4415-b2f4-c2d76adbd072/r )

If I remember correctly, when I tried the same config last week, it didn't show the looping - https://eu.smith.langchain.com/public/55eb198c-4bb5-49f0-be02-898a56f6e365/r

[heorhiikalaichev]

@eyurtsev I have the same issue as @ashburncrash.
BTW, @ashburncrash do you have any progress here?

[ashburncrash]

@heorhiikalaichev Nope - I removed it from using for now. But I'd like to utilise this functionality otherwise I would have to do some extensive manual cleaning on my side 🥲

[ms-tungtt]

I have the same issue, I have to set add_handoff_back_messages to True.
This is my code:

Supervisor:

# supervisor.py

from src.agent.my_agent import graph as my_agent

workflow = create_supervisor(
        [my_agent],
        model=model,
        supervisor_name="supervisor",
        config_schema=ConfigSchema,
        state_schema=MyState,
        prompt=prompt,
        add_handoff_messages=False
    )

My agent:

# my_agent.py

async def agent_factory(state, config: RunnableConfig):    
    mcp_client = get_mcp_client(config)
    
    mcp_tools = await mcp_client.get_tools()

    tools = [
        *mcp_tools,
    ]

    return create_react_agent(
        model=model,
        tools=tools,
        prompt=prompt,
        checkpointer=True,
    )

workflow = StateGraph(state_schema=MyState, config_schema=ConfigSchema)

workflow.add_node("my_agent", agent_factory)
workflow.set_entry_point("my_agent")
workflow.add_edge("my_agent", END)

graph = workflow.compile(name="my_agent")

Is it because I created another StateGraph, cause I need to get RunnableConfig?

[chirico85]

Have the same issue with langgraph-supervisor 0.0.27