Issue with “Using with a custom agent” example when using models that support multi-tool calling (e.g., gpt-4o-mini)

[WorksOnMyVM]

Hi everyone,

I’ve encountered a problem when running the “Using with a custom agent” example in LangGraph with a model that supports multi-tool calling, such as gpt-4o-mini.

For the full code and setup, you can check this Gist:

https://gist.github.com/Makaveli-x/51845482edb6a591fa965d9453aee6a5

Problem

The example breaks when the model calls both the add tool and the make_handoff_tool (e.g., transfer_to_multiplication_expert) in a single turn. Before the handoff to multiplication_expert, the ToolMessage returned from the add function is not added to the state, so when control is passed to multiplication_expert, the corresponding ToolMessage for the add call is missing.

This leads to the following error:

Every tool call (LLM requesting to call a tool) in the message history MUST have a corresponding ToolMessage (result of a tool invocation to return to the LLM) - this is required by most LLM providers.

Root Cause

In the call_tools implementation, when the model makes multiple tool calls in a single turn (e.g., both add and a handoff tool), the function returns two Command objects:

1. A Command that performs the handoff to another agent and updates the state with the new ToolMessage from the handoff tool:

Command(
    goto=agent_name,
    graph=Command.PARENT,
    update={"messages": state["messages"] + [tool_message]},
)

2. A Command that updates the state with the ToolMessage resulting from the add tool call:

Command(update={"messages": [tool_response]})

The issue is that the Command responsible for transferring to multiplication_expert does not include the result of the add tool call. So when the receiving agent resumes, the model is missing the ToolMessage for the add tool call, resulting in the error above.

Have you run into a similar issue when using multi-tool-capable models with agent handoffs? If so, how did you solve it?

Would love to hear your thoughts or if there’s a recommended fix!

[jamesrappazzo]

👋 I'm not sure there's a really stellar solution here. The agent might be assuming that "Always do your portion of calculation before the handoff." means to have that be the first entry in a multiple-tool call situation.

A few things I can think of:

1. Tell each agent it's not allowed to call the handoff tool with any other tools.
2. When the handoff tool is called, check if the last message has a tool call without a corresponding tool message. If it does, raise an exception and let langgraph return the exception to the calling node.

[Divyanshu1kumar]

any updates on this? langgraph sends an exception, but what is the solution to handle this?