Python: [Bug]: Structured Output multiple .run invocations with different schemas fails AzureAIClient

[savvaki]

Description

If you use the same client and call .run twice, each with a different schema in the runtime options, the second call returns None.

Code Sample

from pydantic import BaseModel
from azure.identity.aio import DefaultAzureCredential
from agent_framework.azure import AzureAIClient
from pydantic import ConfigDict

class City(BaseModel):
    name: str
    weather: int
    model_config = ConfigDict(extra="forbid")

class WeatherCity(BaseModel):
    san_fran: City
    new_york: City 
    model_config = ConfigDict(extra="forbid")
    
class Country(BaseModel):
    name: str
    weather: int
    model_config = ConfigDict(extra="forbid")
    
class WeatherCountry(BaseModel):
    america: Country
    united_kingdom: Country
    model_config = ConfigDict(extra="forbid")

    
PROJECT_URL = "" 
DEPLOYMENT = "gpt-4.1"

async with (
        DefaultAzureCredential() as credential,
        AzureAIClient(credential=credential, project_endpoint=PROJECT_URL, model_deployment_name=DEPLOYMENT) as client,
        client.as_agent(name="test") as agent
    ):
    result1 = await agent.run("Make up the weather", options={"response_format": WeatherCity})
    print("First Result")
    print(result1.value)
    result2 = await agent.run("Make up the weather", options={"response_format": WeatherCountry})
    print("Second Result")
    print(result2.value)

Error Messages / Stack Traces

First Result
san_fran=City(name='San Francisco', weather=65) new_york=City(name='New York', weather=78)
Second Result
None

Package Versions

agent-framework-azure-ai: 1.0.0b260116, agent-framework-core: 1.0.0b260116

Python Version

3.13.7

Additional Context

No response

[dmytrostruk]

@savvaki Thanks for reporting this issue.

We will probably need to update the documentation to be more specific, but this is not allowed by the service.
Please see this comment: #3230 (comment)

[savvaki]

Thanks for the quick response @dmytrostruk.

Do you have any suggestions on how to achieve this behavior in a state isolated workflow? My executor manages the agent and I need the structured output schema of the internal self.agent to change dynamically in the handler.

[eavanvalkenburg]

Unfortunately we cannot change this, and the design of the agents in foundry has this fixed, so I would suggest also raising a feedback through the foundry portal and in the meantime you could write a small wrapper that creates a new instance of the azure ai client per invocation which would then also create a new agent (version) in Foundry, but that is def not ideal.