Cleaning up state-isolated agents in workflows

[savvaki]

According to this documentation, the best way to run multiple instances of a workflow in parallel is to use factory functions. The code I am referring to is here:

def create_writer_agent() -> ChatAgent:
    """Factory function to create a Writer agent."""
    return AzureOpenAIChatClient(credential=AzureCliCredential()).create_agent(
        instructions=(
            "You are an excellent content writer. You create new content and edit contents based on the feedback."
        ),
        name="writer_agent",
    )

def create_reviewer_agent() -> ChatAgent:
    """Factory function to create a Reviewer agent."""
    return AzureOpenAIChatClient(credential=AzureCliCredential()).create_agent(
        instructions=(
            "You are an excellent content reviewer."
            "Provide actionable feedback to the writer about the provided content."
            "Provide the feedback in the most concise manner possible."
        ),
        name="reviewer_agent",
    )

builder = WorkflowBuilder()
# Register agent factory functions with the workflow builder
builder.register_agent(factory_func=create_writer_agent, name="writer_agent")
builder.register_agent(factory_func=create_reviewer_agent, name="reviewer_agent")
# Add edges using registered factory function names
builder.add_edge("writer_agent", "reviewer_agent")
builder.set_start_executor("writer_agent")

# Each workflow instance created from the builder will have its own agent
# instances and agent threads
workflow = builder.build()

The Problem:

As seen in the image below, the agents are populating in the foundry but are never removed.

Questions:

1. What is the best practice for cleaning up these agents?
2. The agents are created in the classic foundry portal (not the new foundry portal). Why is this the case?

[Evan Mattson (moonbox3)]

Hi savvaki, thanks for the question. Is there perhaps some other code you're running in conjunction with this code you shared? The current code sample is showing the use of AzureOpenAIChatClient. This is a stateless client that hits an Azure OpenAI resource, and it does not create a server-side agent definition. For agents to be showing up in the classic Foundry portal, one needs to be using the AzureAIAgentClient (versus the AzureAIClient which would show up in the new Foundry portal as a v2 agent) or an Azure OpenAI Assistant which is of type AzureOpenAIAssistantsClient.

With that said, the AzureAIAgentClient has an attribute should_cleanup_agent which defaults to True, so it should be deleting the agent after it's the program ends. Can you let us know which version of agent-framework you're using pip show agent-framework.

[pythonic568]

Evan Mattson (@moonbox3) Thank you for these details. You are correct, the OpenAI client is stateless. I was actually using the AzureAIAgentClient sorry for the confusion. Thank you for the clarity on the different clients for Foundry v1 and v2.

Since register_agent does not allow a generator to be passed, it is difficult to manage the clean up code. There is no way for a .close() to be called when the workflow is complete since the factory manages the creation process. Consider the below creation code for example which is found here:

async def create_azure_ai_agent() -> tuple[Callable[..., Awaitable[Any]], Callable[[], Awaitable[None]]]:
    """Helper method to create an Azure AI agent factory and a close function.

    This makes sure the async context managers are properly handled.
    """
    stack = AsyncExitStack()
    cred = await stack.enter_async_context(AzureCliCredential())

    client = await stack.enter_async_context(AzureAIAgentClient(async_credential=cred))

    async def agent(**kwargs: Any) -> Any:
        return await stack.enter_async_context(client.create_agent(**kwargs))

    async def close() -> None:
        await stack.aclose()

    return agent, close

[savvaki]

Evan Mattson (@moonbox3) I have noticed that using the new client (AzureAIClient) does not clean up the v2 agents in the new Foundry portal.

P.S. (the reply above is me from another account)

This is the code to reproduce:

import asyncio
from azure.identity.aio import DefaultAzureCredential
from agent_framework.azure import AzureAIClient
from agent_framework import ChatAgent, WorkflowBuilder
PROJ_URL = "project url here"
DEPLOYMENT = "deployment here"

def create_writer_agent(client: AzureAIClient) -> ChatAgent:
    """Factory function to create a Writer agent."""
    return client.create_agent(
        instructions=(
            "You are an excellent content writer. You create new content and edit contents based on the feedback."
        ),
        name="writer-agent",
    )

def create_reviewer_agent(client: AzureAIClient) -> ChatAgent:
    """Factory function to create a Reviewer agent."""
    return client.create_agent(
    instructions=(
        "You are an excellent content reviewer."
        "Provide actionable feedback to the writer about the provided content."
        "Provide the feedback in the most concise manner possible."
    ),
    name="reviewer-agent",
)

async def run():
    async with DefaultAzureCredential() as credential:
        async with AzureAIClient(credential=credential, project_endpoint=PROJ_URL, model_deployment_name=DEPLOYMENT) as client:
            async with AzureAIClient(credential=credential, project_endpoint=PROJ_URL, model_deployment_name=DEPLOYMENT) as client2:
                print(f"Building")
                builder = WorkflowBuilder()
                # Register agent factory functions with the workflow builder
                builder.register_agent(factory_func=lambda: create_writer_agent(client), name="writer_agent")
                builder.register_agent(factory_func=lambda: create_reviewer_agent(client2), name="reviewer_agent")
                # Add edges using registered factory function names
                builder.add_edge("writer_agent", "reviewer_agent")
                builder.set_start_executor("writer_agent")
                w1 = builder.build()
                w2 = builder.build()
                result1 = await w1.run("Write me a one line poem")            
                result2 = await w2.run("Write me a one line resignation")

asyncio.run(run())