How to add runtime configuration to the prebuilt.create_react_agent graph?

[markns]

Hey, I'm just getting started with langgraph and am using the prebuilt create_react_agent helper function. I'd love to add runtime configuration so I can play with parameters using a streamlit app, but I can't figure out how to do that. Is the best option currently to reimplement the helper from scratch? Thanks :)

[vbarda]

hi @markns -- could you say more about the parameters you're trying to customize? but overall yes, implementing react agent from scratch in this case might be the best option at the moment

[markns]

hey @vbarda I wanted to be able to swap the model out at runtime, as shown in a few examples. I followed the "from scratch" example, and things are working nicely now.
I would also like to be able to set some model params such as temperature too. I guess this should be possible in the model.invoke call?

    state: AgentState,
    config: RunnableConfig,
):
    model_name = config["configurable"].get("model", "openai-gpt-4o-mini")
    model = models[model_name]

    model = model.bind_tools(tools)
    # this is similar to customizing the create_react_agent with state_modifier, but is a lot more flexible
    system_prompt = SystemMessage(instructions)
    response = model.invoke([system_prompt] + state["messages"], config)
    # We return a list, because this will get added to the existing list
    return {"messages": [response]}
    ```

[Arokamals]

I found this resource which might be a help to you.
Here's the link: How to access the RunnableConfig from a tool

[johschmidt42]

That shows how to access the configurable values within the tools! But you need add configurable fields to your runnable first! There is some documentation regarding that for langchain. But for langgraph, you need to provide the config_schema parameter in StateGraph.

[austinmw]

Hey @vbarda , is it still the best option to create react agent from scratch instead of using prebuilt if you want to be able to change model ID from frontend?

[johschmidt42]

Use the config_schema parameter in StateGraph.