Agents#

Assistant Agent#

AssistantAgent is a built-in agent that uses a language model and has the ability to use tools.

from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.messages import StructuredMessage
from autogen_agentchat.ui import Console
from autogen_ext.models.openai import OpenAIChatCompletionClient

# Define a tool that searches the web for information.
# For simplicity, we will use a mock function here that returns a static string.
async def web_search(query: str) -> str:
    """Find information on the web"""
    return "AutoGen is a programming framework for building multi-agent applications."


# Create an agent that uses the OpenAI GPT-4o model.
model_client = OpenAIChatCompletionClient(
    model="gpt-4.1-nano",
    # api_key="YOUR_API_KEY",
)
agent = AssistantAgent(
    name="assistant",
    model_client=model_client,
    tools=[web_search],
    system_message="Use tools to solve tasks.",
)

Getting Result#

We can use the run() method to get the agent run on a given task.

# Use asyncio.run(agent.run(...)) when running in a script.
result = await agent.run(task="Find information on AutoGen")
print(result.messages)

[TextMessage(source='user', models_usage=None, metadata={}, content='Find information on AutoGen', type='TextMessage'), ToolCallRequestEvent(source='assistant', models_usage=RequestUsage(prompt_tokens=61, completion_tokens=16), metadata={}, content=[FunctionCall(id='call_703i17OLXfztkuioUbkESnea', arguments='{"query":"AutoGen"}', name='web_search')], type='ToolCallRequestEvent'), ToolCallExecutionEvent(source='assistant', models_usage=None, metadata={}, content=[FunctionExecutionResult(content='AutoGen is a programming framework for building multi-agent applications.', name='web_search', call_id='call_703i17OLXfztkuioUbkESnea', is_error=False)], type='ToolCallExecutionEvent'), ToolCallSummaryMessage(source='assistant', models_usage=None, metadata={}, content='AutoGen is a programming framework for building multi-agent applications.', type='ToolCallSummaryMessage')]

The call to the run() method returns a TaskResult with the list of messages in the messages attribute, which stores the agent’s “thought process” as well as the final response.

Multi-Modal Input#

The AssistantAgent can handle multi-modal input by providing the input as a MultiModalMessage.

from io import BytesIO

import PIL
import requests
from autogen_agentchat.messages import MultiModalMessage
from autogen_core import Image

# Create a multi-modal message with random image and text.
pil_image = PIL.Image.open(BytesIO(requests.get("https://picsum.photos/300/200").content))
img = Image(pil_image)
multi_modal_message = MultiModalMessage(content=["Can you describe the content of this image?", img], source="user")
img

# Use asyncio.run(...) when running in a script.
result = await agent.run(task=multi_modal_message)
print(result.messages[-1].content)  # type: ignore

The image depicts a scenic mountain landscape under a clear blue sky. There are several rugged mountain peaks in the background, with some clouds scattered across the sky. In the valley below, there is a body of water, possibly a lake or river, surrounded by greenery. The overall scene conveys a sense of natural beauty and tranquility.

Streaming Messages#

We can also stream each message as it is generated by the agent by using the run_stream() method, and use Console to print the messages as they appear to the console.

async def assistant_run_stream() -> None:
    # Option 1: read each message from the stream (as shown in the previous example).
    # async for message in agent.run_stream(task="Find information on AutoGen"):
    #     print(message)

    # Option 2: use Console to print all messages as they appear.
    await Console(
        agent.run_stream(task="Find information on AutoGen"),
        output_stats=True,  # Enable stats printing.
    )


# Use asyncio.run(assistant_run_stream()) when running in a script.
await assistant_run_stream()

---------- TextMessage (user) ----------
Find information on AutoGen
---------- ToolCallRequestEvent (assistant) ----------
[FunctionCall(id='call_HOTRhOzXCBm0zSqZCFbHD7YP', arguments='{"query":"AutoGen"}', name='web_search')]
[Prompt tokens: 61, Completion tokens: 16]
---------- ToolCallExecutionEvent (assistant) ----------
[FunctionExecutionResult(content='AutoGen is a programming framework for building multi-agent applications.', name='web_search', call_id='call_HOTRhOzXCBm0zSqZCFbHD7YP', is_error=False)]
---------- ToolCallSummaryMessage (assistant) ----------
AutoGen is a programming framework for building multi-agent applications.
---------- Summary ----------
Number of messages: 4
Finish reason: None
Total prompt tokens: 61
Total completion tokens: 16
Duration: 0.52 seconds

The run_stream() method returns an asynchronous generator that yields each message generated by the agent, followed by a TaskResult as the last item.

From the messages, you can observe that the assistant agent utilized the web_search tool to gather information and responded based on the search results.

Using Tools and Workbench#

Large Language Models (LLMs) are typically limited to generating text or code responses. However, many complex tasks benefit from the ability to use external tools that perform specific actions, such as fetching data from APIs or databases.

To address this limitation, modern LLMs can now accept a list of available tool schemas (descriptions of tools and their arguments) and generate a tool call message. This capability is known as Tool Calling or Function Calling and is becoming a popular pattern in building intelligent agent-based applications. Refer to the documentation from OpenAI and Anthropic for more information about tool calling in LLMs.

In AgentChat, the AssistantAgent can use tools to perform specific actions. The web_search tool is one such tool that allows the assistant agent to search the web for information. A single custom tool can be a Python function or a subclass of the BaseTool.

On the other hand, a Workbench is a collection of tools that share state and resources.

By default, when AssistantAgent executes a tool, it will return the tool’s output as a string in ToolCallSummaryMessage in its response. If your tool does not return a well-formed string in natural language, you can add a reflection step to have the model summarize the tool’s output, by setting the reflect_on_tool_use=True parameter in the AssistantAgent constructor.

from autogen_core.tools import FunctionTool


# Define a tool using a Python function.
async def web_search_func(query: str) -> str:
    """Find information on the web"""
    return "AutoGen is a programming framework for building multi-agent applications."


# This step is automatically performed inside the AssistantAgent if the tool is a Python function.
web_search_function_tool = FunctionTool(web_search_func, description="Find information on the web")
# The schema is provided to the model during AssistantAgent's on_messages call.
web_search_function_tool.schema

{'name': 'web_search_func',
 'description': 'Find information on the web',
 'parameters': {'type': 'object',
  'properties': {'query': {'description': 'query',
    'title': 'Query',
    'type': 'string'}},
  'required': ['query'],
  'additionalProperties': False},
 'strict': False}

from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.messages import TextMessage
from autogen_ext.models.openai import OpenAIChatCompletionClient
from autogen_ext.tools.mcp import McpWorkbench, StdioServerParams

# Get the fetch tool from mcp-server-fetch.
fetch_mcp_server = StdioServerParams(command="uvx", args=["mcp-server-fetch"])

# Create an MCP workbench which provides a session to the mcp server.
async with McpWorkbench(fetch_mcp_server) as workbench:  # type: ignore
    # Create an agent that can use the fetch tool.
    model_client = OpenAIChatCompletionClient(model="gpt-4.1-nano")
    fetch_agent = AssistantAgent(
        name="fetcher", model_client=model_client, workbench=workbench, reflect_on_tool_use=True
    )

    # Let the agent fetch the content of a URL and summarize it.
    result = await fetch_agent.run(task="Summarize the content of https://en.wikipedia.org/wiki/Seattle")
    assert isinstance(result.messages[-1], TextMessage)
    print(result.messages[-1].content)

    # Close the connection to the model client.
    await model_client.close()

Seattle is a major city located in the state of Washington, United States. It was founded on November 13, 1851, and incorporated as a town on January 14, 1865, and later as a city on December 2, 1869. The city is named after Chief Seattle. It covers an area of approximately 142 square miles, with a population of around 737,000 as of the 2020 Census, and an estimated 755,078 residents in 2023. Seattle is known by nicknames such as The Emerald City, Jet City, and Rain City, and has mottos including The City of Flowers and The City of Goodwill. The city operates under a mayor–council government system, with Bruce Harrell serving as mayor. Key landmarks include the Space Needle, Pike Place Market, Amazon Spheres, and the Seattle Great Wheel. It is situated on the U.S. West Coast, with a diverse urban and metropolitan area that extends to a population of over 4 million in the greater metropolitan region.

model_client_no_parallel_tool_call = OpenAIChatCompletionClient(
    model="gpt-4o",
    parallel_tool_calls=False,  # type: ignore
)
agent_no_parallel_tool_call = AssistantAgent(
    name="assistant",
    model_client=model_client_no_parallel_tool_call,
    tools=[web_search],
    system_message="Use tools to solve tasks.",
)

Running an Agent in a Loop#

The AssistantAgent executes one step at a time: one model call, followed by one tool call (or parallel tool calls), and then an optional reflection.

To run it in a loop, for example, running it until it stops producing tool calls, please refer to Single-Agent Team.

Structured Output#

Structured output allows models to return structured JSON text with pre-defined schema provided by the application. Different from JSON-mode, the schema can be provided as a Pydantic BaseModel class, which can also be used to validate the output.

Once you specify the base model class in the output_content_type parameter of the AssistantAgent constructor, the agent will respond with a StructuredMessage whose content’s type is the type of the base model class.

This way, you can integrate agent’s response directly into your application and use the model’s output as a structured object.

Structured output is also useful for incorporating Chain-of-Thought reasoning in the agent’s responses. See the example below for how to use structured output with the assistant agent.

from typing import Literal

from pydantic import BaseModel


# The response format for the agent as a Pydantic base model.
class AgentResponse(BaseModel):
    thoughts: str
    response: Literal["happy", "sad", "neutral"]


# Create an agent that uses the OpenAI GPT-4o model.
model_client = OpenAIChatCompletionClient(model="gpt-4o")
agent = AssistantAgent(
    "assistant",
    model_client=model_client,
    system_message="Categorize the input as happy, sad, or neutral following the JSON format.",
    # Define the output content type of the agent.
    output_content_type=AgentResponse,
)

result = await Console(agent.run_stream(task="I am happy."))

# Check the last message in the result, validate its type, and print the thoughts and response.
assert isinstance(result.messages[-1], StructuredMessage)
assert isinstance(result.messages[-1].content, AgentResponse)
print("Thought: ", result.messages[-1].content.thoughts)
print("Response: ", result.messages[-1].content.response)
await model_client.close()

---------- user ----------
I am happy.

---------- assistant ----------
{
  "thoughts": "The user explicitly states they are happy.",
  "response": "happy"
}
Thought:  The user explicitly states they are happy.
Response:  happy

Streaming Tokens#

You can stream the tokens generated by the model client by setting model_client_stream=True. This will cause the agent to yield ModelClientStreamingChunkEvent messages in run_stream().

The underlying model API must support streaming tokens for this to work. Please check with your model provider to see if this is supported.

model_client = OpenAIChatCompletionClient(model="gpt-4o")

streaming_assistant = AssistantAgent(
    name="assistant",
    model_client=model_client,
    system_message="You are a helpful assistant.",
    model_client_stream=True,  # Enable streaming tokens.
)

# Use an async function and asyncio.run() in a script.
async for message in streaming_assistant.run_stream(task="Name two cities in South America"):  # type: ignore
    print(message)

source='user' models_usage=None metadata={} content='Name two cities in South America' type='TextMessage'
source='assistant' models_usage=None metadata={} content='Two' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' cities' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' in' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' South' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' America' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' are' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' Buenos' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' Aires' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' in' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' Argentina' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' and' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' São' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' Paulo' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' in' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' Brazil' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content='.' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=RequestUsage(prompt_tokens=0, completion_tokens=0) metadata={} content='Two cities in South America are Buenos Aires in Argentina and São Paulo in Brazil.' type='TextMessage'
messages=[TextMessage(source='user', models_usage=None, metadata={}, content='Name two cities in South America', type='TextMessage'), TextMessage(source='assistant', models_usage=RequestUsage(prompt_tokens=0, completion_tokens=0), metadata={}, content='Two cities in South America are Buenos Aires in Argentina and São Paulo in Brazil.', type='TextMessage')] stop_reason=None

You can see the streaming chunks in the output above. The chunks are generated by the model client and are yielded by the agent as they are received. The final response, the concatenation of all the chunks, is yielded right after the last chunk.

Using Model Context#

AssistantAgent has a model_context parameter that can be used to pass in a ChatCompletionContext object. This allows the agent to use different model contexts, such as BufferedChatCompletionContext to limit the context sent to the model.

By default, AssistantAgent uses the UnboundedChatCompletionContext which sends the full conversation history to the model. To limit the context to the last n messages, you can use the BufferedChatCompletionContext. To limit the context by token count, you can use the TokenLimitedChatCompletionContext.

from autogen_core.model_context import BufferedChatCompletionContext

# Create an agent that uses only the last 5 messages in the context to generate responses.
agent = AssistantAgent(
    name="assistant",
    model_client=model_client,
    tools=[web_search],
    system_message="Use tools to solve tasks.",
    model_context=BufferedChatCompletionContext(buffer_size=5),  # Only use the last 5 messages in the context.
)

Other Preset Agents#

The following preset agents are available:

• UserProxyAgent: An agent that takes user input returns it as responses.

• CodeExecutorAgent: An agent that can execute code.

• OpenAIAssistantAgent: An agent that is backed by an OpenAI Assistant, with ability to use custom tools.

• MultimodalWebSurfer: A multi-modal agent that can search the web and visit web pages for information.

• FileSurfer: An agent that can search and browse local files for information.

• VideoSurfer: An agent that can watch videos for information.

Next Step#

Having explored the usage of the AssistantAgent, we can now proceed to the next section to learn about the teams feature in AgentChat.