Agent does not pass the tool result to LLM for inference

[ranpariyachetan]

Hi,

I am learning to create a chat bot using AssistantAgent with a FunctionTool to get weather information.

async def get_weather_response(location: str) -> dict:
    """
    Get the weather response for a specific location.
    """
    async with python_weather.Client(unit=python_weather.METRIC) as client:
        weather = await client.get(location)    
        weather_report = {}
        weather_report["location"] = location
        weather_report["temperature"] = weather.temperature
        weather_report["description"] = weather.description
        weather_report["humidity"] = weather.humidity
        weather_report["wind_speed"] = weather.wind_speed
        return weather_report

model_client = OpenAIChatCompletionClient(model="gpt-4.1")

agent =  AssistantAgent(
    name="chatbot",
    model_client=model_client,
    tools=[FunctionTool(get_weather_details, description="Get weather details for a specific location.")],
)

async def main():
    user_input = "How's weather in Singapore?"
    response = await agent.run(
        task=user_input
    )
    print(response.messages[-1].content)

asyncio.run(main()) 

The above code prints JSON returned from the get_weather_response tool as it it.

{
   "location":"Singapore",
   "temperature":30,
   "description":"Partly cloudy",
   "humidity":70,
   "wind_speed":16
}

It looks like the Agent is not sending the tool result to the LLM for inference. It just returns the tool result.
Is this expected? I am not able to find any way to enforce this except to handle the tool call messages explicitly and run another task of handling json output via LLM on Agent.

I am also interested to know if this is a conscious design choice by Microsoft. Other Agent frameworks such as LangGraph and Amazon Strands do not have this limitation.

[ekzhu]

use reflect_on_tool_use=True to enable reflection: https://microsoft.github.io/autogen/stable/user-guide/agentchat-user-guide/tutorial/agents.html#using-tools-and-workbench

I am also interested to know if this is a conscious design choice by Microsoft. Other Agent frameworks such as LangGraph and Amazon Strands do not have this limitation.

Yes. It is because in one-shot scenarios, the tool itself should provide the final result rather than another reflection step, which only adds more token usage. The tool itself should return well-formated output that can be consumed by application.

If you are iterating on tool then you should also set max_tool_iterations to enable iterative tool use.