Python: AzureAIClient behavior -- Only first agent was created/registered in AI Foundry

[chengyu-liu-cs]

Hi,

I am using this new agent client for Foundry Agent V2 APIs. But the behavior was a bit different as I expected.

When I create two agents with the same instance of AzureAIClient, only one agent (ie., the first agent) was created in Foundry and versioned. The second one can be used, but not registered in Foundry (e.g., Example 01).

# Example 1
    credential = AzureCliCredential()
    client =AzureAIClient(async_credential=credential,
                                project_endpoint=os.environ["AZURE_AI_PROJECT_ENDPOINT"],
                                model_deployment_name=os.environ["AZURE_AI_MODEL_DEPLOYMENT_NAME"])

    writer = client.create_agent(
        instructions=("You are a concise copywriter. Provide a single, punchy marketing sentence based on the prompt."),
        name="writer",
    )

    reviewer = client.create_agent(
        instructions=("You are a thoughtful reviewer. Give brief feedback on the previous assistant message."),
        name="reviewer",
    )

In order to register both in Foundry, I needed to create two instances of AzureAIClient (e.g., Example 01). Is this expected behavior?

# Example 2
    writer = AzureAIClient(async_credential=credential,
                                project_endpoint=os.environ["AZURE_AI_PROJECT_ENDPOINT"],
                                model_deployment_name=os.environ["AZURE_AI_MODEL_DEPLOYMENT_NAME"]).create_agent(
                                                                                                            instructions=("You are a concise copywriter. Provide a single, punchy marketing sentence based on the prompt."),
                                                                                                            name="writer",
                                                                                                        )

    reviewer = AzureAIClient(async_credential=credential,
                                project_endpoint=os.environ["AZURE_AI_PROJECT_ENDPOINT"],
                                model_deployment_name=os.environ["AZURE_AI_MODEL_DEPLOYMENT_NAME"]).create_agent(
                                                                                                        instructions=("You are a thoughtful reviewer. Give brief feedback on the previous assistant message."),
                                                                                                        name="reviewer",

[dmytrostruk]

@chengyuliu-msft Thanks for reporting this issue.
At the moment, method create_agent doesn't actually create the agent remotely - it just gets the ChatAgent instance. The agent will be created remotely during run/run_stream operation (in case if existing agent_name and agent_version were not provided).

In the future, most probably we will update create_agent method and rename it for more clarity.

[chengyu-liu-cs]

Thanks for the insights @dmytrostruk.

I had run operations for each agent. Did not include in my example. Here is my example where the second agent was working, but no agent not created in Foundry. Is it a bug or expected behavior?

    credential = AzureCliCredential()
    client =AzureAIClient(async_credential=credential,
                                project_endpoint=os.environ["AZURE_AI_PROJECT_ENDPOINT"],
                                model_deployment_name=os.environ["AZURE_AI_MODEL_DEPLOYMENT_NAME"])

    client_project = AIProjectClient(endpoint=os.environ["AZURE_AI_PROJECT_ENDPOINT"], credential=credential)

    logging.info("Testing OpenAI client access from project client...")
    client_openai = client_project.get_openai_client()
    conversation = await client_openai.conversations.create()
    conversation_id = conversation.id
    print(conversation)
    logging.info("Successfully created conversation with ID: %s", conversation_id)


    writer = client.create_agent(
        instructions=("You are a helpful assistant to help users make notes accurately and rewrite sentences based on users spoken transcripts. "
                "Please fix the grammar errors in the user-provided sentence and make it more readable. "
                "If they appear in the sentence and are misspelled, please fix them. "),
        name="agentv2-test-aiagentclient-writer",
    )
    # Test basic query
    query = "How are you?"
    print(f"User: {query}")
    result = await writer.run(query)
    print(f"Writer Agent: {result}\n")

    reviewer = client.create_agent(
        instructions=("You are a thoughtful reviewer. Give brief feedback on the previous assistant message."),
        name="agentv2-test-aiagentclient-reviewer",
    )

    # Test basic query
    query = "How are you?"
    print(f"User: {query}")
    result = await reviewer.run(query)
    print(f"Reviewer Agent: {result}\n")

[ManniArora]

+1, Running into the same issue - Have a workflow with two Foundry agents, but only one gets created in Foundry, and both runs show up under that single agent.