Production Error with CopilotKit + LangGraph Agent

[edgy basalt]

Hey everyone, I ran into a weird issue when deploying a CopilotKit-based agent to production. Everything works fine locally, but in prod, I’m seeing this error when trying to execute the agent:

TypeError: 'NoneType' object is not a mapping
(config = {**default_config, **(self.graph.config or {}), **config})
It’s coming from the _stream_events function inside copilotkit/langgraph_agent.py.

Local environment: fine
Production: crashes during execution after calling /copilotkit/agents/execute

Here’s what I suspect might be the issue:

• self.graph.config is None in production.

• Possibly due to the config not being initialized properly.

Or something’s missing in the deployment setup that’s present locally.

For backend, I’m using FastAPI to self-host the LangGraph agent with the add_fastapi_endpoint tool to serve the CopilotKit frontend. In front-end, I'm using CopilotKit Runtime.

For production, I'm using AWS ECS to deploy the backend API. I can confirm the copilotkit endpoint gets hit when interacting through the frontend.

Would appreciate any help, insights, or debugging tips. I’ve already done a bit of digging but haven’t found a solid fix yet.

Thanks in advance! 🙏

[manic zodiac]

Hello @edgy basalt
Can you share some code snippets, specifically on the runtime implementation, so I can try to debug it?

In the meantime, you can also compare your code to these examples to see if you have missed anything.

https://github.com/CopilotKit/CopilotKit/tree/main/examples/saas-dynamic-dashboards

[edgy basalt]

Thanks so much for the quick reply, @manic zodiac 🙏

Sure! I’ve attached a few key files that might help pinpoint the issue:

graph.py: sets up and compiles the LangGraph agent.

main.py: FastAPI backend that mounts CopilotKit onto an existing API. I’ve tried applying configs for both LangGraph and CopilotKit here, but the issue persists.

route.ts: the frontend endpoint that uses CopilotKit Runtime to call the backend.

Dockerfile: defines the backend container setup.

Here’s the open-source repo for reference — it mirrors the same architecture and setup:
🔗 https://github.com/al-mz/insight-copilot

For debugging, I even swapped out my graph with the joke-teller agent from CopilotKit examples, deployed it to ECS — and still got the exact same error. So I suspect the root cause might be somewhere in the backend API setup, Dockerfile, or the way the runtime is configured.

Really appreciate you taking the time to look into it!

[manic zodiac]

Thanks for the details. Let me look into this

[slender kraken]

I've seen this in past. it mainly revolves around checkpoints and passing config to langgraph.

# Compile the builder into an executable graph
memory = MemorySaver()
graph = builder.compile(name="powersim_agent")  # ❌ Missing checkpointer

To:

# Compile the builder into an executable graph
memory = MemorySaver()
graph = builder.compile(checkpointer=memory, name="powersim_agent")  # ✅ Fixed

This is the exact pattern that works in all the CopilotKit examples:

from langgraph.checkpoint.memory import MemorySaver
memory = MemorySaver()
graph = workflow.compile(checkpointer=memory)

When you don't provide a checkpointer, LangGraph doesn't initialize the graph's config, leading to the 'NoneType' object is not a mapping error in the CopilotKit runtime.

Alternative if you don't want persistence, you can also just ensure the graph has a config:

graph = builder.compile(name="powersim_agent")
# Ensure config exists for CopilotKit compatibility
if not hasattr(graph, 'config') or graph.config is None:
    graph.config = {}

But the checkpointer approach is cleaner and gives you conversation persistence.

[halcyon basin]

@edgy basalt i faced same issue, u can by pass it temporarily until there is a fix by doing this:

    def execute(self, *args, **kwargs):
        logger.info("Executing custom LangGraphAgent")
        # over ride config parameter

        if "config" not in kwargs or kwargs["config"] is None:
            kwargs["config"] = {}

        return super().execute(*args, **kwargs)```

[edgy basalt]

Thanks @slender kraken , but as you can see in the graph.py attached, the checkpointer has already been included:

memory = MemorySaver()
graph = builder.compile(name="powersim_agent", checkpointer=memory)

@halcyon basin Thanks, I'll give it a try

[edgy basalt]

Hey @halcyon basin ! Your workaround solution worked very well. Thank you very much 🙌

[fallen idol]

Hey @edgy basalt
I will issue a fix for this today which will be available in our python sdk 0.1.47

[fallen idol]

released and available

[tribal bay]

Hey all, looks like we got a few resolutions here. Any lingering issues?