import asyncio
import json
import logging
import warnings
from typing import (
Any,
AsyncGenerator,
Awaitable,
Callable,
Dict,
List,
Mapping,
Optional,
Sequence,
Tuple,
Union,
)
from autogen_core import CancellationToken, Component, ComponentModel, FunctionCall
from autogen_core.memory import Memory
from autogen_core.model_context import (
ChatCompletionContext,
UnboundedChatCompletionContext,
)
from autogen_core.models import (
AssistantMessage,
ChatCompletionClient,
CreateResult,
FunctionExecutionResult,
FunctionExecutionResultMessage,
LLMMessage,
ModelFamily,
SystemMessage,
)
from autogen_core.tools import BaseTool, FunctionTool, StaticWorkbench, Workbench
from pydantic import BaseModel
from typing_extensions import Self
from .. import EVENT_LOGGER_NAME
from ..base import Handoff as HandoffBase
from ..base import Response
from ..messages import (
BaseAgentEvent,
BaseChatMessage,
HandoffMessage,
MemoryQueryEvent,
ModelClientStreamingChunkEvent,
StructuredMessage,
StructuredMessageFactory,
TextMessage,
ThoughtEvent,
ToolCallExecutionEvent,
ToolCallRequestEvent,
ToolCallSummaryMessage,
)
from ..state import AssistantAgentState
from ..utils import remove_images
from ._base_chat_agent import BaseChatAgent
event_logger = logging.getLogger(EVENT_LOGGER_NAME)
class AssistantAgentConfig(BaseModel):
"""助手代理的声明式配置。"""
name: str
model_client: ComponentModel
tools: List[ComponentModel] | None = None
workbench: ComponentModel | None = None
handoffs: List[HandoffBase | str] | None = None
model_context: ComponentModel | None = None
memory: List[ComponentModel] | None = None
description: str
system_message: str | None = None
model_client_stream: bool = False
reflect_on_tool_use: bool
tool_call_summary_format: str
metadata: Dict[str, str] | None = None
structured_message_factory: ComponentModel | None = None
[文档]
class AssistantAgent(BaseChatAgent, Component[AssistantAgentConfig]):
"""An agent that provides assistance with tool use.
The :meth:`on_messages` returns a :class:`~autogen_agentchat.base.Response`
in which :attr:`~autogen_agentchat.base.Response.chat_message` is the final
response message.
The :meth:`on_messages_stream` creates an async generator that produces
the inner messages as they are created, and the :class:`~autogen_agentchat.base.Response`
object as the last item before closing the generator.
The :meth:`BaseChatAgent.run` method returns a :class:`~autogen_agentchat.base.TaskResult`
containing the messages produced by the agent. In the list of messages,
:attr:`~autogen_agentchat.base.TaskResult.messages`,
the last message is the final response message.
The :meth:`BaseChatAgent.run_stream` method creates an async generator that produces
the inner messages as they are created, and the :class:`~autogen_agentchat.base.TaskResult`
object as the last item before closing the generator.
.. attention::
The caller must only pass the new messages to the agent on each call
to the :meth:`on_messages`, :meth:`on_messages_stream`, :meth:`BaseChatAgent.run`,
or :meth:`BaseChatAgent.run_stream` methods.
The agent maintains its state between calls to these methods.
Do not pass the entire conversation history to the agent on each call.
.. warning::
The assistant agent is not thread-safe or coroutine-safe.
It should not be shared between multiple tasks or coroutines, and it should
not call its methods concurrently.
The following diagram shows how the assistant agent works:
.. image:: ../../images/assistant-agent.svg
**Structured output:**
If the `output_content_type` is set, the agent will respond with a :class:`~autogen_agentchat.messages.StructuredMessage`
instead of a :class:`~autogen_agentchat.messages.TextMessage` in the final response by default.
.. note::
Currently, setting `output_content_type` prevents the agent from being
able to call `load_component` and `dum_component` methods for serializable
configuration. This will be fixed soon in the future.
**Tool call behavior:**
* If the model returns no tool call, then the response is immediately returned as a :class:`~autogen_agentchat.messages.TextMessage` or a :class:`~autogen_agentchat.messages.StructuredMessage` (when using structured output) in :attr:`~autogen_agentchat.base.Response.chat_message`.
* When the model returns tool calls, they will be executed right away:
- When `reflect_on_tool_use` is False, the tool call results are returned as a :class:`~autogen_agentchat.messages.ToolCallSummaryMessage` in :attr:`~autogen_agentchat.base.Response.chat_message`. You can customise the summary with either a static format string (`tool_call_summary_format`) **or** a callable (`tool_call_summary_formatter`); the callable is evaluated once per tool call.
- When `reflect_on_tool_use` is True, the another model inference is made using the tool calls and results, and final response is returned as a :class:`~autogen_agentchat.messages.TextMessage` or a :class:`~autogen_agentchat.messages.StructuredMessage` (when using structured output) in :attr:`~autogen_agentchat.base.Response.chat_message`.
- `reflect_on_tool_use` is set to `True` by default when `output_content_type` is set.
- `reflect_on_tool_use` is set to `False` by default when `output_content_type` is not set.
* If the model returns multiple tool calls, they will be executed concurrently. To disable parallel tool calls you need to configure the model client. For example, set `parallel_tool_calls=False` for :class:`~autogen_ext.models.openai.OpenAIChatCompletionClient` and :class:`~autogen_ext.models.openai.AzureOpenAIChatCompletionClient`.
.. tip::
By default, the tool call results are returned as the response when tool
calls are made, so pay close attention to how the tools’ return values
are formatted—especially if another agent expects a specific schema.
* Use **`tool_call_summary_format`** for a simple static template.
* Use **`tool_call_summary_formatter`** for full programmatic control
(e.g., “hide large success payloads, show full details on error”).
*Note*: `tool_call_summary_formatter` is **not serializable** and will
be ignored when an agent is loaded from, or exported to, YAML/JSON
configuration files.
**Hand off behavior:**
* If a handoff is triggered, a :class:`~autogen_agentchat.messages.HandoffMessage` will be returned in :attr:`~autogen_agentchat.base.Response.chat_message`.
* If there are tool calls, they will also be executed right away before returning the handoff.
* The tool calls and results are passed to the target agent through :attr:`~autogen_agentchat.messages.HandoffMessage.context`.
.. note::
If multiple handoffs are detected, only the first handoff is executed.
To avoid this, disable parallel tool calls in the model client configuration.
**Limit context size sent to the model:**
You can limit the number of messages sent to the model by setting
the `model_context` parameter to a :class:`~autogen_core.model_context.BufferedChatCompletionContext`.
This will limit the number of recent messages sent to the model and can be useful
when the model has a limit on the number of tokens it can process.
Another option is to use a :class:`~autogen_core.model_context.TokenLimitedChatCompletionContext`
which will limit the number of tokens sent to the model.
You can also create your own model context by subclassing
:class:`~autogen_core.model_context.ChatCompletionContext`.
**Streaming mode:**
The assistant agent can be used in streaming mode by setting `model_client_stream=True`.
In this mode, the :meth:`on_messages_stream` and :meth:`BaseChatAgent.run_stream` methods will also yield
:class:`~autogen_agentchat.messages.ModelClientStreamingChunkEvent`
messages as the model client produces chunks of response.
The chunk messages will not be included in the final response's inner messages.
Args:
name (str): The name of the agent.
model_client (ChatCompletionClient): The model client to use for inference.
tools (List[BaseTool[Any, Any] | Callable[..., Any] | Callable[..., Awaitable[Any]]] | None, optional): The tools to register with the agent.
workbench (Workbench | None, optional): The workbench to use for the agent.
Tools cannot be used when workbench is set and vice versa.
handoffs (List[HandoffBase | str] | None, optional): The handoff configurations for the agent,
allowing it to transfer to other agents by responding with a :class:`HandoffMessage`.
The transfer is only executed when the team is in :class:`~autogen_agentchat.teams.Swarm`.
If a handoff is a string, it should represent the target agent's name.
model_context (ChatCompletionContext | None, optional): The model context for storing and retrieving :class:`~autogen_core.models.LLMMessage`. It can be preloaded with initial messages. The initial messages will be cleared when the agent is reset.
description (str, optional): The description of the agent.
system_message (str, optional): The system message for the model. If provided, it will be prepended to the messages in the model context when making an inference. Set to `None` to disable.
model_client_stream (bool, optional): If `True`, the model client will be used in streaming mode.
:meth:`on_messages_stream` and :meth:`BaseChatAgent.run_stream` methods will also yield :class:`~autogen_agentchat.messages.ModelClientStreamingChunkEvent`
messages as the model client produces chunks of response. Defaults to `False`.
reflect_on_tool_use (bool, optional): If `True`, the agent will make another model inference using the tool call and result
to generate a response. If `False`, the tool call result will be returned as the response. By default, if `output_content_type` is set, this will be `True`;
if `output_content_type` is not set, this will be `False`.
output_content_type (type[BaseModel] | None, optional): The output content type for :class:`~autogen_agentchat.messages.StructuredMessage` response as a Pydantic model.
This will be used with the model client to generate structured output.
If this is set, the agent will respond with a :class:`~autogen_agentchat.messages.StructuredMessage` instead of a :class:`~autogen_agentchat.messages.TextMessage`
in the final response, unless `reflect_on_tool_use` is `False` and a tool call is made.
output_content_type_format (str | None, optional): (Experimental) The format string used for the content of a :class:`~autogen_agentchat.messages.StructuredMessage` response.
tool_call_summary_format (str, optional): Static format string applied to each tool call result when composing the :class:`~autogen_agentchat.messages.ToolCallSummaryMessage`.
Defaults to ``"{result}"``. Ignored if `tool_call_summary_formatter` is provided. When `reflect_on_tool_use` is ``False``, the summaries for all tool
calls are concatenated with a newline ('\\n') and returned as the response. Placeholders available in the template:
`{tool_name}`, `{arguments}`, `{result}`, `{is_error}`.
tool_call_summary_formatter (Callable[[FunctionCall, FunctionExecutionResult], str] | None, optional):
Callable that receives the ``FunctionCall`` and its ``FunctionExecutionResult`` and returns the summary string.
Overrides `tool_call_summary_format` when supplied and allows conditional logic — for example, emitting static string like
``"Tool FooBar executed successfully."`` on success and a full payload (including all passed arguments etc.) only on failure.
**Limitation**: The callable is *not serializable*; values provided via YAML/JSON configs are ignored.
.. note::
`tool_call_summary_formatter` is intended for in-code use only. It cannot currently be saved or restored via
configuration files.
memory (Sequence[Memory] | None, optional): The memory store to use for the agent. Defaults to `None`.
metadata (Dict[str, str] | None, optional): Optional metadata for tracking.
Raises:
ValueError: If tool names are not unique.
ValueError: If handoff names are not unique.
ValueError: If handoff names are not unique from tool names.
ValueError: If maximum number of tool iterations is less than 1.
Examples:
**Example 1: basic agent**
The following example demonstrates how to create an assistant agent with
a model client and generate a response to a simple task.
.. code-block:: python
import asyncio
from autogen_ext.models.openai import OpenAIChatCompletionClient
from autogen_agentchat.agents import AssistantAgent
async def main() -> None:
model_client = OpenAIChatCompletionClient(
model="gpt-4o",
# api_key = "your_openai_api_key"
)
agent = AssistantAgent(name="assistant", model_client=model_client)
result = await agent.run(task="Name two cities in North America.")
print(result)
asyncio.run(main())
**Example 2: model client token streaming**
This example demonstrates how to create an assistant agent with
a model client and generate a token stream by setting `model_client_stream=True`.
.. code-block:: python
import asyncio
from autogen_ext.models.openai import OpenAIChatCompletionClient
from autogen_agentchat.agents import AssistantAgent
async def main() -> None:
model_client = OpenAIChatCompletionClient(
model="gpt-4o",
# api_key = "your_openai_api_key"
)
agent = AssistantAgent(
name="assistant",
model_client=model_client,
model_client_stream=True,
)
stream = agent.run_stream(task="Name two cities in North America.")
async for message in stream:
print(message)
asyncio.run(main())
.. code-block:: text
source='user' models_usage=None metadata={} content='Name two cities in North 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=' North' 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=' New' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' York' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' City' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' and' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' Toronto' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content='.' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content=' TERMIN' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=None metadata={} content='ATE' type='ModelClientStreamingChunkEvent'
source='assistant' models_usage=RequestUsage(prompt_tokens=0, completion_tokens=0) metadata={} content='Two cities in North America are New York City and Toronto. TERMINATE' type='TextMessage'
messages=[TextMessage(source='user', models_usage=None, metadata={}, content='Name two cities in North America.', type='TextMessage'), TextMessage(source='assistant', models_usage=RequestUsage(prompt_tokens=0, completion_tokens=0), metadata={}, content='Two cities in North America are New York City and Toronto. TERMINATE', type='TextMessage')] stop_reason=None
**Example 3: agent with tools**
The following example demonstrates how to create an assistant agent with
a model client and a tool, generate a stream of messages for a task, and
print the messages to the console using :class:`~autogen_agentchat.ui.Console`.
The tool is a simple function that returns the current time.
Under the hood, the function is wrapped in a :class:`~autogen_core.tools.FunctionTool`
and used with the agent's model client. The doc string of the function
is used as the tool description, the function name is used as the tool name,
and the function signature including the type hints is used as the tool arguments.
.. code-block:: python
import asyncio
from autogen_ext.models.openai import OpenAIChatCompletionClient
from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.ui import Console
async def get_current_time() -> str:
return "The current time is 12:00 PM."
async def main() -> None:
model_client = OpenAIChatCompletionClient(
model="gpt-4o",
# api_key = "your_openai_api_key"
)
agent = AssistantAgent(name="assistant", model_client=model_client, tools=[get_current_time])
await Console(agent.run_stream(task="What is the current time?"))
asyncio.run(main())
**Example 4: agent with Model-Context Protocol (MCP) workbench**
The following example demonstrates how to create an assistant agent with
a model client and an :class:`~autogen_ext.tools.mcp.McpWorkbench` for
interacting with a Model-Context Protocol (MCP) server.
.. code-block:: python
import asyncio
from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.ui import Console
from autogen_ext.models.openai import OpenAIChatCompletionClient
from autogen_ext.tools.mcp import StdioServerParams, McpWorkbench
async def main() -> None:
params = StdioServerParams(
command="uvx",
args=["mcp-server-fetch"],
read_timeout_seconds=60,
)
# You can also use `start()` and `stop()` to manage the session.
async with McpWorkbench(server_params=params) as workbench:
model_client = OpenAIChatCompletionClient(model="gpt-4.1-nano")
assistant = AssistantAgent(
name="Assistant",
model_client=model_client,
workbench=workbench,
reflect_on_tool_use=True,
)
await Console(
assistant.run_stream(task="Go to https://github.com/microsoft/autogen and tell me what you see.")
)
asyncio.run(main())
**Example 5: agent with structured output and tool**
The following example demonstrates how to create an assistant agent with
a model client configured to use structured output and a tool.
Note that you need to use :class:`~autogen_core.tools.FunctionTool` to create the tool
and the `strict=True` is required for structured output mode.
Because the model is configured to use structured output, the output
reflection response will be a JSON formatted string.
.. code-block:: python
import asyncio
from typing import Literal
from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.ui import Console
from autogen_core.tools import FunctionTool
from autogen_ext.models.openai import OpenAIChatCompletionClient
from pydantic import BaseModel
# Define the structured output format.
class AgentResponse(BaseModel):
thoughts: str
response: Literal["happy", "sad", "neutral"]
# Define the function to be called as a tool.
def sentiment_analysis(text: str) -> str:
\"\"\"Given a text, return the sentiment.\"\"\"
return "happy" if "happy" in text else "sad" if "sad" in text else "neutral"
# Create a FunctionTool instance with `strict=True`,
# which is required for structured output mode.
tool = FunctionTool(sentiment_analysis, description="Sentiment Analysis", strict=True)
# Create an OpenAIChatCompletionClient instance that supports structured output.
model_client = OpenAIChatCompletionClient(
model="gpt-4o-mini",
)
# Create an AssistantAgent instance that uses the tool and model client.
agent = AssistantAgent(
name="assistant",
model_client=model_client,
tools=[tool],
system_message="Use the tool to analyze sentiment.",
output_content_type=AgentResponse,
)
async def main() -> None:
stream = agent.run_stream(task="I am happy today!")
await Console(stream)
asyncio.run(main())
.. code-block:: text
---------- assistant ----------
[FunctionCall(id='call_tIZjAVyKEDuijbBwLY6RHV2p', arguments='{"text":"I am happy today!"}', name='sentiment_analysis')]
---------- assistant ----------
[FunctionExecutionResult(content='happy', call_id='call_tIZjAVyKEDuijbBwLY6RHV2p', is_error=False)]
---------- assistant ----------
{"thoughts":"The user expresses a clear positive emotion by stating they are happy today, suggesting an upbeat mood.","response":"happy"}
**Example 6: agent with bounded model context**
The following example shows how to use a
:class:`~autogen_core.model_context.BufferedChatCompletionContext`
that only keeps the last 2 messages (1 user + 1 assistant).
Bounded model context is useful when the model has a limit on the
number of tokens it can process.
.. code-block:: python
import asyncio
from autogen_agentchat.agents import AssistantAgent
from autogen_core.model_context import BufferedChatCompletionContext
from autogen_ext.models.openai import OpenAIChatCompletionClient
async def main() -> None:
# Create a model client.
model_client = OpenAIChatCompletionClient(
model="gpt-4o-mini",
# api_key = "your_openai_api_key"
)
# Create a model context that only keeps the last 2 messages (1 user + 1 assistant).
model_context = BufferedChatCompletionContext(buffer_size=2)
# Create an AssistantAgent instance with the model client and context.
agent = AssistantAgent(
name="assistant",
model_client=model_client,
model_context=model_context,
system_message="You are a helpful assistant.",
)
result = await agent.run(task="Name two cities in North America.")
print(result.messages[-1].content) # type: ignore
result = await agent.run(task="My favorite color is blue.")
print(result.messages[-1].content) # type: ignore
result = await agent.run(task="Did I ask you any question?")
print(result.messages[-1].content) # type: ignore
asyncio.run(main())
.. code-block:: text
Two cities in North America are New York City and Toronto.
That's great! Blue is often associated with calmness and serenity. Do you have a specific shade of blue that you like, or any particular reason why it's your favorite?
No, you didn't ask a question. I apologize for any misunderstanding. If you have something specific you'd like to discuss or ask, feel free to let me know!
**Example 7: agent with memory**
The following example shows how to use a list-based memory with the assistant agent.
The memory is preloaded with some initial content.
Under the hood, the memory is used to update the model context
before making an inference, using the :meth:`~autogen_core.memory.Memory.update_context` method.
.. code-block:: python
import asyncio
from autogen_agentchat.agents import AssistantAgent
from autogen_core.memory import ListMemory, MemoryContent
from autogen_ext.models.openai import OpenAIChatCompletionClient
async def main() -> None:
# Create a model client.
model_client = OpenAIChatCompletionClient(
model="gpt-4o-mini",
# api_key = "your_openai_api_key"
)
# Create a list-based memory with some initial content.
memory = ListMemory()
await memory.add(MemoryContent(content="User likes pizza.", mime_type="text/plain"))
await memory.add(MemoryContent(content="User dislikes cheese.", mime_type="text/plain"))
# Create an AssistantAgent instance with the model client and memory.
agent = AssistantAgent(
name="assistant",
model_client=model_client,
memory=[memory],
system_message="You are a helpful assistant.",
)
result = await agent.run(task="What is a good dinner idea?")
print(result.messages[-1].content) # type: ignore
asyncio.run(main())
.. code-block:: text
How about making a delicious pizza without cheese? You can create a flavorful veggie pizza with a variety of toppings. Here's a quick idea:
**Veggie Tomato Sauce Pizza**
- Start with a pizza crust (store-bought or homemade).
- Spread a layer of marinara or tomato sauce evenly over the crust.
- Top with your favorite vegetables like bell peppers, mushrooms, onions, olives, and spinach.
- Add some protein if you’d like, such as grilled chicken or pepperoni (ensure it's cheese-free).
- Sprinkle with herbs like oregano and basil, and maybe a drizzle of olive oil.
- Bake according to the crust instructions until the edges are golden and the veggies are cooked.
Serve it with a side salad or some garlic bread to complete the meal! Enjoy your dinner!
**Example 8: agent with `o1-mini`**
The following example shows how to use `o1-mini` model with the assistant agent.
.. code-block:: python
import asyncio
from autogen_ext.models.openai import OpenAIChatCompletionClient
from autogen_agentchat.agents import AssistantAgent
async def main() -> None:
model_client = OpenAIChatCompletionClient(
model="o1-mini",
# api_key = "your_openai_api_key"
)
# The system message is not supported by the o1 series model.
agent = AssistantAgent(name="assistant", model_client=model_client, system_message=None)
result = await agent.run(task="What is the capital of France?")
print(result.messages[-1].content) # type: ignore
asyncio.run(main())
.. note::
The `o1-preview` and `o1-mini` models do not support system message and function calling.
So the `system_message` should be set to `None` and the `tools` and `handoffs` should not be set.
See `o1 beta limitations <https://platform.openai.com/docs/guides/reasoning#beta-limitations>`_ for more details.
**Example 9: agent using reasoning model with custom model context.**
The following example shows how to use a reasoning model (DeepSeek R1) with the assistant agent.
The model context is used to filter out the thought field from the assistant message.
.. code-block:: python
import asyncio
from typing import List
from autogen_agentchat.agents import AssistantAgent
from autogen_core.model_context import UnboundedChatCompletionContext
from autogen_core.models import AssistantMessage, LLMMessage, ModelFamily
from autogen_ext.models.ollama import OllamaChatCompletionClient
class ReasoningModelContext(UnboundedChatCompletionContext):
\"\"\"A model context for reasoning models.\"\"\"
async def get_messages(self) -> List[LLMMessage]:
messages = await super().get_messages()
# Filter out thought field from AssistantMessage.
messages_out: List[LLMMessage] = []
for message in messages:
if isinstance(message, AssistantMessage):
message.thought = None
messages_out.append(message)
return messages_out
# Create an instance of the model client for DeepSeek R1 hosted locally on Ollama.
model_client = OllamaChatCompletionClient(
model="deepseek-r1:8b",
model_info={
"vision": False,
"function_calling": False,
"json_output": False,
"family": ModelFamily.R1,
"structured_output": True,
},
)
agent = AssistantAgent(
"reasoning_agent",
model_client=model_client,
model_context=ReasoningModelContext(), # Use the custom model context.
)
async def run_reasoning_agent() -> None:
result = await agent.run(task="What is the capital of France?")
print(result)
asyncio.run(run_reasoning_agent())
"""
component_config_schema = AssistantAgentConfig
component_provider_override = "autogen_agentchat.agents.AssistantAgent"
def __init__(
self,
name: str,
model_client: ChatCompletionClient,
*,
tools: List[BaseTool[Any, Any] | Callable[..., Any] | Callable[..., Awaitable[Any]]] | None = None,
workbench: Workbench | None = None,
handoffs: List[HandoffBase | str] | None = None,
model_context: ChatCompletionContext | None = None,
description: str = "An agent that provides assistance with ability to use tools.",
system_message: (
str | None
) = "You are a helpful AI assistant. Solve tasks using your tools. Reply with TERMINATE when the task has been completed.",
model_client_stream: bool = False,
reflect_on_tool_use: bool | None = None,
tool_call_summary_format: str = "{result}",
tool_call_summary_formatter: Callable[[FunctionCall, FunctionExecutionResult], str] | None = None,
output_content_type: type[BaseModel] | None = None,
output_content_type_format: str | None = None,
memory: Sequence[Memory] | None = None,
metadata: Dict[str, str] | None = None,
):
super().__init__(name=name, description=description)
self._metadata = metadata or {}
self._model_client = model_client
self._model_client_stream = model_client_stream
self._output_content_type: type[BaseModel] | None = output_content_type
self._output_content_type_format = output_content_type_format
self._structured_message_factory: StructuredMessageFactory | None = None
if output_content_type is not None:
self._structured_message_factory = StructuredMessageFactory(
input_model=output_content_type, format_string=output_content_type_format
)
self._memory = None
if memory is not None:
if isinstance(memory, list):
self._memory = memory
else:
raise TypeError(f"Expected Memory, List[Memory], or None, got {type(memory)}")
self._system_messages: List[SystemMessage] = []
if system_message is None:
self._system_messages = []
else:
self._system_messages = [SystemMessage(content=system_message)]
self._tools: List[BaseTool[Any, Any]] = []
if tools is not None:
if model_client.model_info["function_calling"] is False:
raise ValueError("The model does not support function calling.")
for tool in tools:
if isinstance(tool, BaseTool):
self._tools.append(tool)
elif callable(tool):
if hasattr(tool, "__doc__") and tool.__doc__ is not None:
description = tool.__doc__
else:
description = ""
self._tools.append(FunctionTool(tool, description=description))
else:
raise ValueError(f"Unsupported tool type: {type(tool)}")
# Check if tool names are unique.
tool_names = [tool.name for tool in self._tools]
if len(tool_names) != len(set(tool_names)):
raise ValueError(f"Tool names must be unique: {tool_names}")
# Handoff tools.
self._handoff_tools: List[BaseTool[Any, Any]] = []
self._handoffs: Dict[str, HandoffBase] = {}
if handoffs is not None:
if model_client.model_info["function_calling"] is False:
raise ValueError("The model does not support function calling, which is needed for handoffs.")
for handoff in handoffs:
if isinstance(handoff, str):
handoff = HandoffBase(target=handoff)
if isinstance(handoff, HandoffBase):
self._handoff_tools.append(handoff.handoff_tool)
self._handoffs[handoff.name] = handoff
else:
raise ValueError(f"Unsupported handoff type: {type(handoff)}")
# Check if handoff tool names are unique.
handoff_tool_names = [tool.name for tool in self._handoff_tools]
if len(handoff_tool_names) != len(set(handoff_tool_names)):
raise ValueError(f"Handoff names must be unique: {handoff_tool_names}")
# Check if handoff tool names not in tool names.
if any(name in tool_names for name in handoff_tool_names):
raise ValueError(
f"Handoff names must be unique from tool names. "
f"Handoff names: {handoff_tool_names}; tool names: {tool_names}"
)
if workbench is not None:
if self._tools:
raise ValueError("Tools cannot be used with a workbench.")
self._workbench = workbench
else:
self._workbench = StaticWorkbench(self._tools)
if model_context is not None:
self._model_context = model_context
else:
self._model_context = UnboundedChatCompletionContext()
if self._output_content_type is not None and reflect_on_tool_use is None:
# If output_content_type is set, we need to reflect on tool use by default.
self._reflect_on_tool_use = True
elif reflect_on_tool_use is None:
self._reflect_on_tool_use = False
else:
self._reflect_on_tool_use = reflect_on_tool_use
if self._reflect_on_tool_use and ModelFamily.is_claude(model_client.model_info["family"]):
warnings.warn(
"Claude models may not work with reflection on tool use because Claude requires that any requests including a previous tool use or tool result must include the original tools definition."
"Consider setting reflect_on_tool_use to False. "
"As an alternative, consider calling the agent in a loop until it stops producing tool calls. "
"See [Single-Agent Team](https://microsoft.github.io/autogen/stable/user-guide/agentchat-user-guide/tutorial/teams.html#single-agent-team) "
"for more details.",
UserWarning,
stacklevel=2,
)
self._tool_call_summary_format = tool_call_summary_format
self._tool_call_summary_formatter = tool_call_summary_formatter
self._is_running = False
@property
def produced_message_types(self) -> Sequence[type[BaseChatMessage]]:
message_types: List[type[BaseChatMessage]] = []
if self._handoffs:
message_types.append(HandoffMessage)
if self._tools:
message_types.append(ToolCallSummaryMessage)
if self._output_content_type:
message_types.append(StructuredMessage[self._output_content_type]) # type: ignore[name-defined]
else:
message_types.append(TextMessage)
return tuple(message_types)
@property
def model_context(self) -> ChatCompletionContext:
"""
代理正在使用的模型上下文。
"""
return self._model_context
[文档]
async def on_messages(self, messages: Sequence[BaseChatMessage], cancellation_token: CancellationToken) -> Response:
async for message in self.on_messages_stream(messages, cancellation_token):
if isinstance(message, Response):
return message
raise AssertionError("The stream should have returned the final result.")
[文档]
async def on_messages_stream(
self, messages: Sequence[BaseChatMessage], cancellation_token: CancellationToken
) -> AsyncGenerator[BaseAgentEvent | BaseChatMessage | Response, None]:
"""
处理传入消息与助手代理交互，并实时生成事件/响应。
"""
# Gather all relevant state here
agent_name = self.name
model_context = self._model_context
memory = self._memory
system_messages = self._system_messages
workbench = self._workbench
handoff_tools = self._handoff_tools
handoffs = self._handoffs
model_client = self._model_client
model_client_stream = self._model_client_stream
reflect_on_tool_use = self._reflect_on_tool_use
tool_call_summary_format = self._tool_call_summary_format
tool_call_summary_formatter = self._tool_call_summary_formatter
output_content_type = self._output_content_type
format_string = self._output_content_type_format
# STEP 1: Add new user/handoff messages to the model context
await self._add_messages_to_context(
model_context=model_context,
messages=messages,
)
# STEP 2: Update model context with any relevant memory
inner_messages: List[BaseAgentEvent | BaseChatMessage] = []
for event_msg in await self._update_model_context_with_memory(
memory=memory,
model_context=model_context,
agent_name=agent_name,
):
inner_messages.append(event_msg)
yield event_msg
# STEP 3: Run the first inference
model_result = None
async for inference_output in self._call_llm(
model_client=model_client,
model_client_stream=model_client_stream,
system_messages=system_messages,
model_context=model_context,
workbench=workbench,
handoff_tools=handoff_tools,
agent_name=agent_name,
cancellation_token=cancellation_token,
output_content_type=output_content_type,
):
if isinstance(inference_output, CreateResult):
model_result = inference_output
else:
# Streaming chunk event
yield inference_output
assert model_result is not None, "No model result was produced."
# --- NEW: If the model produced a hidden "thought," yield it as an event ---
if model_result.thought:
thought_event = ThoughtEvent(content=model_result.thought, source=agent_name)
yield thought_event
inner_messages.append(thought_event)
# Add the assistant message to the model context (including thought if present)
await model_context.add_message(
AssistantMessage(
content=model_result.content,
source=agent_name,
thought=getattr(model_result, "thought", None),
)
)
# STEP 4: Process the model output
async for output_event in self._process_model_result(
model_result=model_result,
inner_messages=inner_messages,
cancellation_token=cancellation_token,
agent_name=agent_name,
system_messages=system_messages,
model_context=model_context,
workbench=workbench,
handoff_tools=handoff_tools,
handoffs=handoffs,
model_client=model_client,
model_client_stream=model_client_stream,
reflect_on_tool_use=reflect_on_tool_use,
tool_call_summary_format=tool_call_summary_format,
tool_call_summary_formatter=tool_call_summary_formatter,
output_content_type=output_content_type,
format_string=format_string,
):
yield output_event
@staticmethod
async def _add_messages_to_context(
model_context: ChatCompletionContext,
messages: Sequence[BaseChatMessage],
) -> None:
"""
将传入消息添加到模型上下文中。
"""
for msg in messages:
if isinstance(msg, HandoffMessage):
for llm_msg in msg.context:
await model_context.add_message(llm_msg)
await model_context.add_message(msg.to_model_message())
@staticmethod
async def _update_model_context_with_memory(
memory: Optional[Sequence[Memory]],
model_context: ChatCompletionContext,
agent_name: str,
) -> List[MemoryQueryEvent]:
"""
如果存在记忆模块，则更新模型上下文并返回产生的事件。
"""
events: List[MemoryQueryEvent] = []
if memory:
for mem in memory:
update_context_result = await mem.update_context(model_context)
if update_context_result and len(update_context_result.memories.results) > 0:
memory_query_event_msg = MemoryQueryEvent(
content=update_context_result.memories.results,
source=agent_name,
)
events.append(memory_query_event_msg)
return events
@classmethod
async def _call_llm(
cls,
model_client: ChatCompletionClient,
model_client_stream: bool,
system_messages: List[SystemMessage],
model_context: ChatCompletionContext,
workbench: Workbench,
handoff_tools: List[BaseTool[Any, Any]],
agent_name: str,
cancellation_token: CancellationToken,
output_content_type: type[BaseModel] | None,
) -> AsyncGenerator[Union[CreateResult, ModelClientStreamingChunkEvent], None]:
"""
执行模型推理并生成流式分块事件或最终的CreateResult。
"""
all_messages = await model_context.get_messages()
llm_messages = cls._get_compatible_context(model_client=model_client, messages=system_messages + all_messages)
tools = (await workbench.list_tools()) + handoff_tools
if model_client_stream:
model_result: Optional[CreateResult] = None
async for chunk in model_client.create_stream(
llm_messages,
tools=tools,
json_output=output_content_type,
cancellation_token=cancellation_token,
):
if isinstance(chunk, CreateResult):
model_result = chunk
elif isinstance(chunk, str):
yield ModelClientStreamingChunkEvent(content=chunk, source=agent_name)
else:
raise RuntimeError(f"Invalid chunk type: {type(chunk)}")
if model_result is None:
raise RuntimeError("No final model result in streaming mode.")
yield model_result
else:
model_result = await model_client.create(
llm_messages,
tools=tools,
cancellation_token=cancellation_token,
json_output=output_content_type,
)
yield model_result
@classmethod
async def _process_model_result(
cls,
model_result: CreateResult,
inner_messages: List[BaseAgentEvent | BaseChatMessage],
cancellation_token: CancellationToken,
agent_name: str,
system_messages: List[SystemMessage],
model_context: ChatCompletionContext,
workbench: Workbench,
handoff_tools: List[BaseTool[Any, Any]],
handoffs: Dict[str, HandoffBase],
model_client: ChatCompletionClient,
model_client_stream: bool,
reflect_on_tool_use: bool,
tool_call_summary_format: str,
tool_call_summary_formatter: Callable[[FunctionCall, FunctionExecutionResult], str] | None,
output_content_type: type[BaseModel] | None,
format_string: str | None = None,
) -> AsyncGenerator[BaseAgentEvent | BaseChatMessage | Response, None]:
"""
处理来自model_result的最终或部分响应，包括工具调用、交接以及必要的反思。
"""
# If direct text response (string)
if isinstance(model_result.content, str):
if output_content_type:
content = output_content_type.model_validate_json(model_result.content)
yield Response(
chat_message=StructuredMessage[output_content_type]( # type: ignore[valid-type]
content=content,
source=agent_name,
models_usage=model_result.usage,
format_string=format_string,
),
inner_messages=inner_messages,
)
else:
yield Response(
chat_message=TextMessage(
content=model_result.content,
source=agent_name,
models_usage=model_result.usage,
),
inner_messages=inner_messages,
)
return
# Otherwise, we have function calls
assert isinstance(model_result.content, list) and all(
isinstance(item, FunctionCall) for item in model_result.content
)
# STEP 4A: Yield ToolCallRequestEvent
tool_call_msg = ToolCallRequestEvent(
content=model_result.content,
source=agent_name,
models_usage=model_result.usage,
)
event_logger.debug(tool_call_msg)
inner_messages.append(tool_call_msg)
yield tool_call_msg
# STEP 4B: Execute tool calls
executed_calls_and_results = await asyncio.gather(
*[
cls._execute_tool_call(
tool_call=call,
workbench=workbench,
handoff_tools=handoff_tools,
agent_name=agent_name,
cancellation_token=cancellation_token,
)
for call in model_result.content
]
)
exec_results = [result for _, result in executed_calls_and_results]
# Yield ToolCallExecutionEvent
tool_call_result_msg = ToolCallExecutionEvent(
content=exec_results,
source=agent_name,
)
event_logger.debug(tool_call_result_msg)
await model_context.add_message(FunctionExecutionResultMessage(content=exec_results))
inner_messages.append(tool_call_result_msg)
yield tool_call_result_msg
# STEP 4C: Check for handoff
handoff_output = cls._check_and_handle_handoff(
model_result=model_result,
executed_calls_and_results=executed_calls_and_results,
inner_messages=inner_messages,
handoffs=handoffs,
agent_name=agent_name,
)
if handoff_output:
yield handoff_output
return
# STEP 4D: Reflect or summarize tool results
if reflect_on_tool_use:
async for reflection_response in cls._reflect_on_tool_use_flow(
system_messages=system_messages,
model_client=model_client,
model_client_stream=model_client_stream,
model_context=model_context,
agent_name=agent_name,
inner_messages=inner_messages,
output_content_type=output_content_type,
):
yield reflection_response
else:
yield cls._summarize_tool_use(
executed_calls_and_results=executed_calls_and_results,
inner_messages=inner_messages,
handoffs=handoffs,
tool_call_summary_format=tool_call_summary_format,
tool_call_summary_formatter=tool_call_summary_formatter,
agent_name=agent_name,
)
@staticmethod
def _check_and_handle_handoff(
model_result: CreateResult,
executed_calls_and_results: List[Tuple[FunctionCall, FunctionExecutionResult]],
inner_messages: List[BaseAgentEvent | BaseChatMessage],
handoffs: Dict[str, HandoffBase],
agent_name: str,
) -> Optional[Response]:
"""
检测交接调用，根据需要生成HandoffMessage并返回Response。
如果存在多个交接，仅使用第一个。
"""
handoff_reqs = [
call for call in model_result.content if isinstance(call, FunctionCall) and call.name in handoffs
]
if len(handoff_reqs) > 0:
# We have at least one handoff function call
selected_handoff = handoffs[handoff_reqs[0].name]
if len(handoff_reqs) > 1:
warnings.warn(
(
f"Multiple handoffs detected. Only the first is executed: "
f"{[handoffs[c.name].name for c in handoff_reqs]}. "
"Disable parallel tool calls in the model client to avoid this warning."
),
stacklevel=2,
)
# Collect normal tool calls (not handoff) into the handoff context
tool_calls: List[FunctionCall] = []
tool_call_results: List[FunctionExecutionResult] = []
# Collect the results returned by handoff_tool. By default, the message attribute will returned.
selected_handoff_message = selected_handoff.message
for exec_call, exec_result in executed_calls_and_results:
if exec_call.name not in handoffs:
tool_calls.append(exec_call)
tool_call_results.append(exec_result)
elif exec_call.name == selected_handoff.name:
selected_handoff_message = exec_result.content
handoff_context: List[LLMMessage] = []
if len(tool_calls) > 0:
# Include the thought in the AssistantMessage if model_result has it
handoff_context.append(
AssistantMessage(
content=tool_calls,
source=agent_name,
thought=getattr(model_result, "thought", None),
)
)
handoff_context.append(FunctionExecutionResultMessage(content=tool_call_results))
elif model_result.thought:
# If no tool calls, but a thought exists, include it in the context
handoff_context.append(
AssistantMessage(
content=model_result.thought,
source=agent_name,
)
)
# Return response for the first handoff
return Response(
chat_message=HandoffMessage(
content=selected_handoff_message,
target=selected_handoff.target,
source=agent_name,
context=handoff_context,
),
inner_messages=inner_messages,
)
return None
@classmethod
async def _reflect_on_tool_use_flow(
cls,
system_messages: List[SystemMessage],
model_client: ChatCompletionClient,
model_client_stream: bool,
model_context: ChatCompletionContext,
agent_name: str,
inner_messages: List[BaseAgentEvent | BaseChatMessage],
output_content_type: type[BaseModel] | None,
) -> AsyncGenerator[Response | ModelClientStreamingChunkEvent | ThoughtEvent, None]:
"""
如果 reflect_on_tool_use=True，我们会基于工具结果进行另一次推理
并生成最终的文本响应（或流式数据块）。
"""
all_messages = system_messages + await model_context.get_messages()
llm_messages = cls._get_compatible_context(model_client=model_client, messages=all_messages)
reflection_result: Optional[CreateResult] = None
if model_client_stream:
async for chunk in model_client.create_stream(
llm_messages,
json_output=output_content_type,
):
if isinstance(chunk, CreateResult):
reflection_result = chunk
elif isinstance(chunk, str):
yield ModelClientStreamingChunkEvent(content=chunk, source=agent_name)
else:
raise RuntimeError(f"Invalid chunk type: {type(chunk)}")
else:
reflection_result = await model_client.create(llm_messages, json_output=output_content_type)
if not reflection_result or not isinstance(reflection_result.content, str):
raise RuntimeError("Reflect on tool use produced no valid text response.")
# --- NEW: If the reflection produced a thought, yield it ---
if reflection_result.thought:
thought_event = ThoughtEvent(content=reflection_result.thought, source=agent_name)
yield thought_event
inner_messages.append(thought_event)
# Add to context (including thought if present)
await model_context.add_message(
AssistantMessage(
content=reflection_result.content,
source=agent_name,
thought=getattr(reflection_result, "thought", None),
)
)
if output_content_type:
content = output_content_type.model_validate_json(reflection_result.content)
yield Response(
chat_message=StructuredMessage[output_content_type]( # type: ignore[valid-type]
content=content,
source=agent_name,
models_usage=reflection_result.usage,
),
inner_messages=inner_messages,
)
else:
yield Response(
chat_message=TextMessage(
content=reflection_result.content,
source=agent_name,
models_usage=reflection_result.usage,
),
inner_messages=inner_messages,
)
@staticmethod
def _summarize_tool_use(
executed_calls_and_results: List[Tuple[FunctionCall, FunctionExecutionResult]],
inner_messages: List[BaseAgentEvent | BaseChatMessage],
handoffs: Dict[str, HandoffBase],
tool_call_summary_format: str,
tool_call_summary_formatter: Callable[[FunctionCall, FunctionExecutionResult], str] | None,
agent_name: str,
) -> Response:
"""
如果 reflect_on_tool_use=False，则创建所有工具调用的摘要消息。
"""
# Filter out calls which were actually handoffs
normal_tool_calls = [(call, result) for call, result in executed_calls_and_results if call.name not in handoffs]
def default_tool_call_summary_formatter(call: FunctionCall, result: FunctionExecutionResult) -> str:
return tool_call_summary_format
summary_formatter = tool_call_summary_formatter or default_tool_call_summary_formatter
tool_call_summaries = [
summary_formatter(call, result).format(
tool_name=call.name,
arguments=call.arguments,
result=result.content,
is_error=result.is_error,
)
for call, result in normal_tool_calls
]
tool_call_summary = "\n".join(tool_call_summaries)
return Response(
chat_message=ToolCallSummaryMessage(
content=tool_call_summary,
source=agent_name,
),
inner_messages=inner_messages,
)
@staticmethod
async def _execute_tool_call(
tool_call: FunctionCall,
workbench: Workbench,
handoff_tools: List[BaseTool[Any, Any]],
agent_name: str,
cancellation_token: CancellationToken,
) -> Tuple[FunctionCall, FunctionExecutionResult]:
"""执行单个工具调用并返回结果。"""
# Load the arguments from the tool call.
try:
arguments = json.loads(tool_call.arguments)
except json.JSONDecodeError as e:
return (
tool_call,
FunctionExecutionResult(
content=f"Error: {e}",
call_id=tool_call.id,
is_error=True,
name=tool_call.name,
),
)
# Check if the tool call is a handoff.
# TODO: consider creating a combined workbench to handle both handoff and normal tools.
for handoff_tool in handoff_tools:
if tool_call.name == handoff_tool.name:
# Run handoff tool call.
result = await handoff_tool.run_json(arguments, cancellation_token)
result_as_str = handoff_tool.return_value_as_string(result)
return (
tool_call,
FunctionExecutionResult(
content=result_as_str,
call_id=tool_call.id,
is_error=False,
name=tool_call.name,
),
)
# Handle normal tool call using workbench.
result = await workbench.call_tool(
name=tool_call.name,
arguments=arguments,
cancellation_token=cancellation_token,
)
return (
tool_call,
FunctionExecutionResult(
content=result.to_text(),
call_id=tool_call.id,
is_error=result.is_error,
name=tool_call.name,
),
)
[文档]
async def on_reset(self, cancellation_token: CancellationToken) -> None:
"""将助手代理重置到初始化状态。"""
await self._model_context.clear()
[文档]
async def save_state(self) -> Mapping[str, Any]:
"""保存助手代理的当前状态。"""
model_context_state = await self._model_context.save_state()
return AssistantAgentState(llm_context=model_context_state).model_dump()
[文档]
async def load_state(self, state: Mapping[str, Any]) -> None:
"""加载助手代理的状态"""
assistant_agent_state = AssistantAgentState.model_validate(state)
# Load the model context state.
await self._model_context.load_state(assistant_agent_state.llm_context)
@staticmethod
def _get_compatible_context(model_client: ChatCompletionClient, messages: List[LLMMessage]) -> Sequence[LLMMessage]:
"""确保消息与底层客户端兼容，必要时移除图像。"""
if model_client.model_info["vision"]:
return messages
else:
return remove_images(messages)
def _to_config(self) -> AssistantAgentConfig:
"""将助手代理转换为声明式配置。"""
return AssistantAgentConfig(
name=self.name,
model_client=self._model_client.dump_component(),
tools=None, # versionchanged:: v0.5.5 Now tools are not serialized, Cause they are part of the workbench.
workbench=self._workbench.dump_component() if self._workbench else None,
handoffs=list(self._handoffs.values()) if self._handoffs else None,
model_context=self._model_context.dump_component(),
memory=[memory.dump_component() for memory in self._memory] if self._memory else None,
description=self.description,
system_message=self._system_messages[0].content
if self._system_messages and isinstance(self._system_messages[0].content, str)
else None,
model_client_stream=self._model_client_stream,
reflect_on_tool_use=self._reflect_on_tool_use,
tool_call_summary_format=self._tool_call_summary_format,
structured_message_factory=self._structured_message_factory.dump_component()
if self._structured_message_factory
else None,
metadata=self._metadata,
)
@classmethod
def _from_config(cls, config: AssistantAgentConfig) -> Self:
"""从声明式配置创建助手代理。"""
if config.structured_message_factory:
structured_message_factory = StructuredMessageFactory.load_component(config.structured_message_factory)
format_string = structured_message_factory.format_string
output_content_type = structured_message_factory.ContentModel
else:
format_string = None
output_content_type = None
return cls(
name=config.name,
model_client=ChatCompletionClient.load_component(config.model_client),
workbench=Workbench.load_component(config.workbench) if config.workbench else None,
handoffs=config.handoffs,
model_context=ChatCompletionContext.load_component(config.model_context) if config.model_context else None,
tools=[BaseTool.load_component(tool) for tool in config.tools] if config.tools else None,
memory=[Memory.load_component(memory) for memory in config.memory] if config.memory else None,
description=config.description,
system_message=config.system_message,
model_client_stream=config.model_client_stream,
reflect_on_tool_use=config.reflect_on_tool_use,
tool_call_summary_format=config.tool_call_summary_format,
output_content_type=output_content_type,
output_content_type_format=format_string,
metadata=config.metadata,
)
