[Python] Langgraph Supervisor SDK not storing states

I am developing a langgraph agent ( Python ) using langgraph SDK,langgraph supervisor. My agent has this structure below

1. 
   Primary Agent ( which will route requests to the appropriate sub-agent )
   
   1. It doesn’t have any tools
2. 
   Order Agent ( which will create / update / save Orders ) - Tools below
   
   1. 
      create_order( source, destination, items ) - This creates an order based on the user's request changes in a dict ( with source, destination, items as keys ) and stores it in shared State
      
      
   2. 
      update_order( source,destination,items) - Does something similar to create_order
      
      
   3. 
      save_order() - Fetch the dict ( Order ) from state and send to API
      

The issue is with the save_order tool. I try to fetch the constructed order from the graph state, but it is always empty. I am updating the Graph state using the Command method in the create & update_order tool.

# AgentStateSchema

from langchain.agents import AgentState

class AgentStateSchema(AgentState):
draft_order: Annotated[dict, update_dicts] = {}


# Tools

from langchain.tools import tool,BaseTool,ToolRuntime

@tool
def create_order(# Some order related fields):
# Redacted the code for brevity
return Command(
update={
"draft_order" : order,
"messages" : [
ToolMessage(content="Created a Draft Order",tool_call_id=tool_call_id)
]
}
)

@tool
def update_order(# Some order related fields):
# Redacted the code for brevity
return Command(
update={
"draft_order" : updated_order,
"messages" : [
ToolMessage(content="Updated the Draft Order",tool_call_id=tool_call_id)
]
}
)

@tool
def save_order(runtime: ToolRuntime):
# Redacted the code for brevity
draft_order = runtime.state.get("draft_order",{}) ##### The draft_order field is always empty


# Order Agent

order_agent = create_agent(
model=ChatOpenAI(model="gpt-4o",temperature=0,max_completion_tokens=None,max_retries=2),
tools=[create_order,update_order,save_order],
system_prompt=prompt.format(),
name="order_agent"
)


# Supervisor

primary_assistant = create_supervisor(
agents=[
order_agent
],
model=primary_assistant_llm,
prompt=primary_assistant_prompt,
output_mode="full_history",
supervisor_name="primary_assistant",
state_schema=AgentStateSchema
).compile(name="primary_assistant")


I tried to make the state to the sub-agent level, but it still gives the same error

Continue reading...