Re: Issue with Multi agent supervisor based agenti...

mark_ott · ‎11-07-2025

Your multi-agent framework has several issues affecting agent iteration control and error handling. Here’s a lined explanation and practical suggestions for each problem:

Agents not stopping at one iteration

Agents consistently reaching the max iteration count (and then the END node) is usually a logic or termination-condition problem in the agent loop—either in your Supervisor logic or in the agent-node response handling. You should check:

The stopping condition after each agent step (i.e., are you checking for a response/event that signals completion, or only waiting for max iterations?)
Whether your agents return a message or status that is supposed to end the iteration. If not, agents will loop until forced to stop.

Suggestion:
Add a proper check after each iteration for a completion signal (“done”, “end”, or a special token/message), and exit early from the loop if found. Ensure that your agent nodes update their state correctly to flag completion.

Coder agent validation error

Error:
validation error for ChatAgentMessage Value error, Either 'content' or 'tool_calls' must be provided.

This error means you’re creating a message dictionary for the agent, but it’s missing both the 'content' and 'tool_calls' fields—which are both required. Most frameworks (like LangChain, UC, or custom ones) check for at least one to exist per message object.

Cause:
Your agent isn’t returning a valid message; either its output is empty, not returning a dict with these keys, or you aren’t formatting the output properly before building the agent message.

Suggestion:
Make sure the Coder agent always returns something like:

python

{
    "role": "assistant",
    "content": response_content,  # This should be a string
    "tool_calls": tool_calls  # Or None if not using tools
}

At least one of content or tool_calls must be not None . If you’re not using tools, ensure content is a non-empty string.

Weather agent error and fix

Previous error:
'PyFuncModel' object has no attribute 'invoke'

You tried to fix this by writing:

python

def agent_node_existing_model(state, agent, name):
    result = agent.predict(state)
    return {
        "messages": [
            {
                "role": "assistant",
                "content": result["messages"][-1].content,
                "name": name,
            }
        ]
    }

Current error:
AttributeError: 'list' object has no attribute 'get'

Cause:
This comes from the format of result. You are calling result["messages"][-1].content, expecting result to be a dict with a "messages" key containing a list of message objects.
But result from agent.predict(state) is apparently a list, not a dict, or its structure is different.

Suggestion:
Print or inspect the exact format of result returned by agent.predict(state). Is it a dict? Or is it a list of dicts? Adjust your code accordingly:

If result is a list of messages:

python

result = agent.predict(state)
last_message = result[-1]  # If result is a list
return {
    "messages": [
        {
            "role": "assistant",
            "content": last_message.get("content", ""),
            "name": name,
        }
    ]
}

If result is a dict:

python

last_message = result["messages"][-1]
return {
    "messages": [
        {
            "role": "assistant",
            "content": last_message.get("content", ""),
            "name": name,
        }
    ]
}

Add checks and debug with print(result) to confirm its structure.

Key Code Patterns for Debugging

Always print or inspect agent outputs before trying to access nested keys.
Make sure message dicts always have valid fields (role, content, and optionally name or tool_calls).
Ensure the loop termination in your Supervisor is tied to agent-completion logic.