8000 Releases · microsoft/autogen · GitHub
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content

Releases: microsoft/autogen

python-v0.5.6

02 May 22:55
@ekzhu ekzhu
python-v0.5.6
880a225
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.
Compare
Choose a tag to compare
python-v0.5.6 Latest
Latest

What's New

GraphFlow: customized workflows using directed graph

Should I say finally? Yes, finally, we have workflows in AutoGen. GraphFlow is a new team class as part of the AgentChat API. One way to think of GraphFlow is that it is a version of SelectorGroupChat but with a directed graph as the selector_func. However, it is actually more powerful, because the abstraction also supports concurrent agents.

Note: GraphFlow is still an experimental API. Watch out for changes in the future releases.

For more details, see our newly added user guide on GraphFlow.

If you are in a hurry, here is an example of creating a fan-out-fan-in workflow:

import asyncio
from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.teams import DiGraphBuilder, GraphFlow
from autogen_agentchat.ui import Console
from autogen_ext.models.openai import OpenAIChatCompletionClient


async def main() -> None:
    # Create an OpenAI model client
    client = OpenAIChatCompletionClient(model="gpt-4.1-nano")

    # Create the writer agent
    writer = AssistantAgent(
        "writer",
        model_client=client,
        system_message="Draft a short paragraph on climate change.",
    )

    # Create two editor agents
    editor1 = AssistantAgent(
        "editor1", model_client=client, system_message="Edit the paragraph for grammar."
    )

    editor2 = AssistantAgent(
        "editor2", model_client=client, system_message="Edit the paragraph for style."
    )

    # Create the final reviewer agent
    final_reviewer = AssistantAgent(
        "final_reviewer",
        model_client=client,
        system_message="Consolidate the grammar and style edits into a final version.",
    )

    # Build the workflow graph
    builder = DiGraphBuilder()
    builder.add_node(writer).add_node(editor1).add_node(editor2).add_node(
        final_reviewer
    )

    # Fan-out from writer to editor1 and editor2
    builder.add_edge(writer, editor1)
    builder.add_edge(writer, editor2)

    # Fan-in both editors into final reviewer
    builder.add_edge(editor1, final_reviewer)
    builder.add_edge(editor2, final_reviewer)

    # Build and validate the graph
    graph = builder.build()

    # Create the flow
    flow = GraphFlow(
        participants=builder.get_participants(),
        graph=graph,
    )

    # Run the workflow
    await Console(flow.run_stream(task="Write a short biography of Steve Jobs."))

asyncio.run(main())

Major thanks to @abhinav-aegis for the initial design and implementation of this amazing feature!

Azure AI Agent Improvement

New Sample

  • A multi-agent PostgreSQL data management example by @mehrsa in #6443

Bug Fixes:

Dev Improvement

Other Python Related Changes

New Contributors

Full Changelog: python-v0.5.5...python-v0.5.6

Contributors

abdomohamed, ekzhu, and 9 other contributors
Assets 2
Loading

python-v0.5.5

25 Apr 23:56
@ekzhu ekzhu
python-v0.5.5
653bcc5
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.
Compare
Choose a tag to compare
python-v0.5.5

What's New

Introduce Workbench

A workbench is a collection of tools that share state and resource. For example, you can now use MCP server through McpWorkbench rather than using tool adapters. This makes it possible to use MCP servers that requires a shared session among the tools (e.g., login session).

Here is an example of using AssistantAgent with GitHub MCP Server.

import asyncio
import os
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 McpWorkbench, StdioServerParams

async def main() -> None:
    model_client = OpenAIChatCompletionClient(model="gpt-4.1-nano")
    server_params = StdioServerParams(
        command="docker",
        args=[
            "run",
            "-i",
            "--rm",
            "-e",
            "GITHUB_PERSONAL_ACCESS_TOKEN",
            "ghcr.io/github/github-mcp-server",
        ],
        env={
            "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_XXXXXXXXXXXXXXXXXXXXXXXXXXXXXX",
        }
    )
    async with McpWorkbench(server_params) as mcp:
        agent = AssistantAgent(
            "github_assistant",
            model_client=model_client,
            workbench=mcp,
            reflect_on_tool_use=True,
            model_client_stream=True,
        )
        await Console(agent.run_stream(task="Is there a repository named Autogen"))
    
asyncio.run(main())

Here is another example showing a web browsing agent using Playwright MCP Server, AssistantAgent and RoundRobinGroupChat.

# First run `npm install -g @playwright/mcp@latest` to install the MCP server.
import asyncio
from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.teams import RoundRobinGroupChat
from autogen_agentchat.conditions import TextMessageTermination
from autogen_agentchat.ui import Console
from autogen_ext.models.openai import OpenAIChatCompletionClient
from autogen_ext.tools.mcp import McpWorkbench, StdioServerParams

async def main() -> None:
    model_client = OpenAIChatCompletionClient(model="gpt-4.1-nano")
    server_params = StdioServerParams(
        command="npx",
        args=[
            "@playwright/mcp@latest",
            "--headless",
        ],
    )
    async with McpWorkbench(server_params) as mcp:
        agent = AssistantAgent(
            "web_browsing_assistant",
            model_client=model_client,
            workbench=mcp,
            model_client_stream=True,
        )
        team = RoundRobinGroupChat(
            [agent],
            termination_condition=TextMessageTermination(source="web_browsing_assistant"),
        )
        await Console(team.run_stream(task="Find out how many contributors for the microsoft/autogen repository"))
    
asyncio.run(main())

Read more:

New Sample: AutoGen and FastAPI with Streaming

  • Add example using autogen-core and FastAPI for handoff multi-agent design pattern with streaming and UI by @amith-ajith in #6391

New Termination Condition: FunctionalTermination

  • Support using a function expression to create a termination condition for teams. by @ekzhu in #6398

Other Python Related Changes

  • update website version by @ekzhu in #6364
  • TEST/change gpt4, gpt4o serise to gpt4.1nano by @SongChiYoung in #6375
  • Remove name field from OpenAI Assistant Message by @ekzhu in #6388
  • Add guide for workbench and mcp & bug fixes for create_mcp_server_session by @ekzhu in #6392
  • TEST: skip when macos+uv and adding uv venv tests by @SongChiYoung in #6387
  • AssistantAgent to support Workbench by @ekzhu in #6393
  • Update agent documentation by @ekzhu in #6394
  • Update version to 0.5.5 by @ekzhu in #6397
  • Update: implement return_value_as_string for McpToolAdapter by @perfogic in #6380
  • [doc] Clarify selector prompt for SelectorGroupChat by @ekzhu in #6399
  • Document custom message types in teams API docs by @ekzhu in #6400

New Contributors

Full Changelog: python-v0.5.4...python-v0.5.5

Contributors

ekzhu, SongChiYoung, and 2 other contributors
Loading

python-v0.5.4

22 Apr 17:51
@ekzhu ekzhu
python-v0.5.4
aad6caa
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.
Compare
Choose a tag to compare
python-v0.5.4

What's New

Agent and Team as Tools

You can use AgentTool and TeamTool to wrap agent and team into tools to be used by other agents.

import asyncio

from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.tools import AgentTool
from autogen_agentchat.ui import Console
from autogen_ext.models.openai import OpenAIChatCompletionClient


async def main() -> None:
    model_client = OpenAIChatCompletionClient(model="gpt-4")
    writer = AssistantAgent(
        name="writer",
        description="A writer agent for generating text.",
        model_client=model_client,
        system_message="Write well.",
    )
    writer_tool = AgentTool(agent=writer)
    assistant = AssistantAgent(
        name="assistant",
        model_client=model_client,
        tools=[writer_tool],
        system_message="You are a helpful assistant.",
    )
    await Console(assistant.run_stream(task="Write a poem about the sea."))


asyncio.run(main())

See AgentChat Tools API for more information.

Azure AI Agent

Introducing adapter for Azure AI Agent, with support for file search, code interpreter, and more. See our Azure AI Agent Extension API.

Docker Jupyter Code Executor

Thinking about sandboxing your local Jupyter execution environment? We just added a new code executor to our family of code executors. See Docker Jupyter Code Executor Extension API.

  • Make Docker Jupyter support to the Version 0.4 as Version 0.2 by @masquerlin in #6231

Canvas Memory

Shared "whiteboard" memory can be useful for agents to collaborate on a common artifact such code, document, or illustration. Canvas Memory is an experimental extension for sharing memory and exposing tools for agents to operate on the shared memory.

New Community Extensions

Updated links to new community extensions. Notably, autogen-contextplus provides advanced model context implementations with ability to automatically summarize, truncate the model context used by agents.

SelectorGroupChat Update

SelectorGroupChat now works with models that only support streaming mode (e.g., QwQ). It can also optionally emit the inner reasoning of the model used in the selector. Set emit_team_events=True and model_client_streaming=True when creating SelectorGroupChat.

  • FEAT: SelectorGroupChat could using stream inner select_prompt by @SongChiYoung in #6286

CodeExecutorAgent Update

CodeExecutorAgent just got another refresh: it now supports max_retries_on_error parameter. You can specify how many times it can retry and self-debug in case there is error in the code execution.

ModelInfo Update

New Sample: AutoGen Core + FastAPI with Streaming

  • Add an example using autogen-core and FastAPI to create streaming responses by @ToryPan in #6335

AGBench Update

Bug Fixes

  • Bugfix: Azure AI Search Tool - fix query type by @jay-thakur in #6331
  • fix: ensure serialized messages are passed to LLMStreamStartEvent by @peterj in #6344
  • fix: ollama fails when tools use optional args by @peterj in #6343
  • Avoid re-registering a message type already registered by @jorge-wonolo in #6354
  • Fix: deserialize model_context in AssistantAgent and SocietyOfMindAgent and CodeExecutorAgent by @SongChiYoung in #6337

What's Changed

  • Update website 0.5.3 by @ekzhu in #6320
  • Update version 0.5.4 by @ekzhu in #6334
  • Generalize Continuous SystemMessage merging via model_info[“multiple_system_messages”] instead of startswith("gemini-") by @SongChiYoung in #6345
  • Add experimental notice to canvas by @ekzhu in #6349
  • Added support for exposing GPUs to docker code executor by @millerh1 in #6339

New Contributors

Full Changelog: python-v0.5.3...python-v0.5.4

Contributors

abdomohamed, ekzhu, and 11 other contributors
Loading

python-v0.5.3

17 Apr 03:36
@ekzhu ekzhu
python-v0.5.3
fb16d5a
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.
Compare
Choose a tag to compare
python-v0.5.3

What's New

CodeExecutorAgent Update

Now the CodeExecutorAgent can generate and execute code in the same invocation. See API doc for examples.

AssistantAgent Improvement

Now AssistantAgent can be serialized when output_content_type is set, thanks @abhinav-aegis's new built-in utility module autogen_core.utils for working with JSON schema.

Team Improvement

Added an optional parameter emit_team_events to configure whether team events like SelectorSpeakerEvent are emitted through run_stream.

  • [FEATURE] Option to emit group chat manager messages in AgentChat by @SongChiYoung in #6303

MCP Improvement

Now mcp_server_tools factory can reuse a shared session. See example of AssistantAgent using Playwright MCP server in the API Doc.

  • Make shared session possible for MCP tool by @ekzhu in #6312

Console Improvement

Bug Fixes

  • Fix: Azure AI Search Tool Client Lifetime Management by @jay-thakur in #6316
  • Make sure thought content is included in handoff context by @ekzhu in #6319

Python Related Changes

New Contributors

Full Changelog: python-v0.5.2...python-v0.5.3

Contributors

ekzhu, jay-thakur, and 7 other contributors
Loading

python-v0.5.2

15 Apr 03:24
@ekzhu ekzhu
python-v0.5.2
3500170
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.
Compare
Choose a tag to compare
python-v0.5.2

Python Related Changes

New Contributors

Full Changelog: python-v0.5.1...python-v0.5.2

Contributors

ekzhu, scovetta, and 13 other contributors
Loading

python-v0.5.1

03 Apr 23:37
@ekzhu ekzhu
python-v0.5.1
47602ea
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.
Compare
Choose a tag to compare
python-v0.5.1

What's New

AgentChat Message Types (Type Hint Changes)

Important

TL;DR: If you are not using custom agents or custom termination conditions, you don't need to change anything.
Otherwise, update AgentEvent to BaseAgentEvent and ChatMessage to BaseChatMessage in your type hints.

This is a breaking change on type hinting only, not on usage.

We updated the message types in AgentChat in this new release.
The purpose of this change is to support custom message types defined by applications.

Previously, message types are fixed and we use the union types ChatMessage and AgentEvent to refer to all the concrete built-in message types.

Now, in the main branch, the message types are organized into hierarchy: existing built-in concrete message types are subclassing either BaseChatMessage and BaseAgentEvent, depending it was part of the ChatMessage or AgentEvent union. We refactored all message handlers on_messages, on_messages_stream, run, run_stream and TerminationCondition to use the base classes in their type hints.

If you are subclassing BaseChatAgent to create your custom agents, or subclassing TerminationCondition to create your custom termination conditions, then you need to rebase the method signatures to use BaseChatMessage and BaseAgentEvent.

If you are using the union types in your existing data structures for serialization and deserialization, then you can keep using those union types to ensure the messages are being handled as concrete types. However, this will not work with custom message types.

Otherwise, your code should just work, as the refactor only makes type hint changes.

This change allows us to support custom message types. For example, we introduced a new message type StructureMessage[T] generic, that can be used to create new message types with a BaseModel content. On-going work is to get AssistantAgent to respond with StructuredMessage[T] where T is the structured output type for the model.

See the API doc on AgentChat message types: https://microsoft.github.io/autogen/stable/reference/python/autogen_agentchat.messages.html

  • Use class hierarchy to organize AgentChat message types and introduce StructuredMessage type by @ekzhu in #5998
  • Rename to use BaseChatMessage and BaseAgentEvent. Bring back union types. by @ekzhu in #6144

Structured Output

We enhanced support for structured output in model clients and agents.

For model clients, use json_output parameter to specify the structured output type
as a Pydantic model. The model client will then return a JSON string
that can be deserialized into the specified Pydantic model.

import asyncio
from typing import Literal

from autogen_core import CancellationToken
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"]

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

 # Generate a response using the tool.
response = await model_client.create(
    messages=[
        SystemMessage(content="Analyze input text sentiment using the tool provided."),
        UserMessage(content="I am happy.", source="user"),
    ],
    json_ouput=AgentResponse,
)

print(response.content)
# Should be a structured output.
# {"thoughts": "The user is happy.", "response": "happy"}

For AssistantAgent, you can set output_content_type to the structured output type. The agent will automatically reflect on the tool call result and generate a StructuredMessage with the output content type.

import asyncio
from typing import Literal

from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.messages import TextMessage
from autogen_agentchat.ui import Console
from autogen_core import CancellationToken
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,
)

stream = agent.on_messages_stream([TextMessage(content="I am happy today!", source="user")], CancellationToken())
await Console(stream)
---------- 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"}

You can also pass a StructuredMessage to the run and run_stream methods of agents and teams as task messages. Agents will automatically deserialize the message to string and place them in their model context. StructuredMessage generated by an agent will also be passed to other agents in the team, and emitted as messages in the output stream.

  • Add structured output to model clients by @ekzhu in #5936
  • Support json schema for response format type in OpenAIChatCompletionClient by @ekzhu in #5988
  • Add output_format to AssistantAgent for structured output by @ekzhu in #6071

Azure AI Search Tool

Added a new tool for agents to perform search using Azure AI Search.

See the documentation for more details.

SelectorGroupChat Improvements

  • Implement 'candidate_func' parameter to filter down the pool of candidates for selection by @Ethan0456 in #5954
  • Add async support for selector_func and candidate_func in SelectorGroupChat by @Ethan0456 in #6068

Code Executors Improvements

  • Add cancellation support to docker executor by @ekzhu in #6027
  • Move start() and stop() as interface methods for CodeExecutor by @ekzhu in #6040
  • Changed Code Executors default directory to temporary directory by @federicovilla55 in #6143

Model Client Improvements

  • Improve documentation around model client and tool and how it works under the hood by @ekzhu in #6050
  • Add support for thought field in AzureAIChatCompletionClient by @jay-thakur in #6062
  • Add a thought process analysis, and add a reasoning field in the ModelClientStreamingChunkEvent to distinguish the thought tokens. by @y26s4824k264 in #5989
  • Add thought field support and fix LLM control parameters for OllamaChatCompletionClient by @jay-thakur in #6126
  • Modular Transformer Pipeline and Fix Gemini/Anthropic Empty Content Handling by @SongChiYoung in #6063
  • Doc/moudulor transform oai by @SongChiYoung in #6149
  • Model family resolution to support non-prefixed names like Mistral by @SongChiYoung in #6158

TokenLimitedChatCompletionContext

Introduce TokenLimitedChatCompletionContext to limit the number of tokens in the context
sent to the model.
This is useful for long-running agents that need to keep a long history of messages in the context.

Bug Fixes

Read more

Contributors

ekzhu, yusufk, and 28 other contributors
Loading

python-v0.4.9.3

29 Mar 04:40
@ekzhu ekzhu
python-v0.4.9.3
e45a157
Compare
Choose a tag to compare
python-v0.4.9.3

Patch Release

This release addresses a bug in MCP Server Tool that causes error when unset tool arguments are set to None and passed on to the server. It also improves the error message from server and adds a default timeout. #6080, #6125

Full Changelog: python-v0.4.9.2...python-v0.4.9.3

Loading

autogenstudio-v0.4.2

17 Mar 18:06
@victordibia victordibia
autogenstudio-v0.4.2
8f8ee04
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.
Compare
Choose a tag to compare
autogenstudio-v0.4.2

What's New

This release makes improvements to AutoGen Studio across multiple areas.

Component Validation and Testing

  • Support Component Validation API in AGS in #5503
  • Test components - #5963
image

In the team builder, all component schemas are automatically validated on save. This way configuration errors (e.g., incorrect provider names) are highlighted early.

In addition, there is a test button for m F438 odel clients where you can verify the correctness of your model configuration. The LLM is given a simple query and the results are shown.

Gallery Improvements

  • Improved editing UI for tools in AGS by in #5539
  • Anthropic support in AGS #5695

You can now modify teams, agents, models, tools, and termination conditions independently in the UI, and only review JSON when needed. The same UI panel for updating components in team builder is also reused in the Gallery. The Gallery in AGS is now persisted in a database, rather than local storage. Anthropic models supported in AGS.
image

Observability - LLMCallEvents

  • Enable LLM Call Observability in AGS #5457

You can now view all LLMCallEvents in AGS. Go to settings (cog icon on lower left) to enable this feature.

Token Streaming

  • Add Token Streaming in AGS in #5659

For better developer experience, the AGS UI will stream tokens as they are generated by an LLM for any agent where stream_model_client is set to true.

UX Improvements - Session Comparison

  • AGS - Test Model Component in UI, Compare Sessions in #5963

It is often valuable, even critical, to have a side-by-side comparison of multiple agent configurations (e.g., using a team of web agents that solve tasks using a browser or agents with web search API tools). You can now do this using the compare button in the playground, which lets you select multiple sessions and interact with them to compare outputs.

Experimental Features

There are a few interesting but early features that ship with this release:

  • Authentication in AGS: You can pass in an authentication configuration YAML file to enable user authentication for AGS. Currently, only GitHub authentication is supported. This lays the foundation for a multi-user environment (#5928) where various users can login and only view their own sessions. More work needs to be done to clarify isolation of resources (e.g., environment variables) and other security considerations.
    See the documentation for more details.
loginags.mov
image
  • Local Python Code Execution Tool: AGS now has early support for a local Python code execution tool. More work is needed to test the underlying agentchat implementation

Other Fixes

  • Fixed issue with using AzureSQL DB as the database engine for AGS
  • Fixed cascading delete issue in AGS (ensure runs are deleted when sessions are deleted) #5804 by @victordibia
  • Fixed termination UI bug #5888
  • Fixed DockerFile for AGS by @gunt3001 #5932

Thanks to @ekzhu , @jackgerrits , @gagb, @usag1e, @dominiclachance , @EItanya and many others for testing and feedback

Contributors

ekzhu, victordibia, and 6 other contributors
Loading

python-v0.4.9.2

14 Mar 19:40
@ekzhu ekzhu
python-v0.4.9.2
4a1660f
Compare
Choose a tag to compare
python-v0.4.9.2

Patch Fixes

  • Fix logging error in SKChatCompletionAdapter #5893
  • Fix missing system message in the model client call during reflect step when reflect_on_tool_use=True #5926 (Bug introduced in v0.4.8)
  • Fixing listing directory error in FileSurfer #5938

Security Fixes

  • Use SecretStr type for model clients' API key. This will ensure the secret is not exported when calling model_client.dump_component().model_dump_json(). #5939 and #5947. This will affect OpenAIChatCompletionClient and AzureOpenAIChatCompletionClient, and AnthropicChatCompletionClient -- the API keys will no longer be exported when you serialize the model clients. It is recommended to use environment-based or token-based authentication rather than passing the API keys around as data in configs.

Full Changelog: python-v0.4.9...python-v0.4.9.2

Loading

python-v0.4.9

12 Mar 07:21
@ekzhu ekzhu
python-v0.4.9
bb8439c
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.
Compare
Choose a tag to compare
python-v0.4.9

What's New

Anthropic Model Client

Native support for Anthropic models. Get your update:
 

pip install -U "autogen-ext[anthropic]"

The new client follows the same interface as OpenAIChatCompletionClient so you can use it directly in your agents and teams.

import asyncio
from autogen_ext.models.anthropic import AnthropicChatCompletionClient
from autogen_core.models import UserMessage


async def main():
    anthropic_client = AnthropicChatCompletionClient(
        model="claude-3-sonnet-20240229",
        api_key="your-api-key",  # Optional if ANTHROPIC_API_KEY is set in environment
    )

    result = await anthropic_client.create([UserMessage(content="What is the capital of France?", source="user")])  # type: ignore
    print(result)


if __name__ == "__main__":
    asyncio.run(main())

You can also load the model client directly from a configuration dictionary:

from autogen_core.models import ChatCompletionClient

config = {
    "provider": "AnthropicChatCompletionClient",
    "config": {"model": "claude-3-sonnet-20240229"},
}

client = ChatCompletionClient.load_component(config)

To use with AssistantAgent and run the agent in a loop to match the behavior of Claude agents, you can use Single-Agent Team.

LlamaCpp Model Client

LlamaCpp is a great project for working with local models. Now we have native support via its official SDK.

pip install -U "autogen-ext[llama-cpp]"

To use a local model file:

import asyncio

from autogen_core.models import UserMessage
from autogen_ext.models.llama_
AF56
cpp import LlamaCppChatCompletionClient


async def main():
    llama_client = LlamaCppChatCompletionClient(model_path="/path/to/your/model.gguf")
    result = await llama_client.create([UserMessage(content="What is the capital of France?", source="user")])
    print(result)


asyncio.run(main())

To use it with a Hugging Face model:

import asyncio

from autogen_core.models import UserMessage
from autogen_ext.models.llama_cpp import LlamaCppChatCompletionClient


async def main():
    llama_client = LlamaCppChatCompletionClient(
        repo_id="unsloth/phi-4-GGUF", filename="phi-4-Q2_K_L.gguf", n_gpu_layers=-1, seed=1337, n_ctx=5000
    )
    result = await llama_client.create([UserMessage(content="What is the capital of France?", source="user")])
    print(result)


asyncio.run(main())

Task-Centric Memory (Experimental)

Task-Centric memory is an experimental module that can give agents the ability to:

  • Accomplish general tasks more effectively by learning quickly and continually beyond context-window limitations.
  • Remember guidance, corrections, plans, and demonstrations provided by users (teachability)
  • Learn through the agent's own experience and adapt quickly to changing circumstances (self-improvement)
  • Avoid repeating mistakes on tasks that are similar to those previously encountered.

For example, you can use Teachability as a memory for AssistantAgent so your agent can learn from user teaching.

from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.ui import Console
from autogen_ext.models.openai import OpenAIChatCompletionClient
from autogen_ext.experimental.task_centric_memory import MemoryController
from autogen_ext.experimental.task_centric_memory.utils import Teachability


async def main():
    # Create a client
    client = OpenAIChatCompletionClient(model="gpt-4o-2024-08-06", )

    # Create an instance of Task-Centric Memory, passing minimal parameters for this simple example
    memory_controller = MemoryController(reset=False, client=client)

    # Wrap the memory controller in a Teachability instance
    teachability = Teachability(memory_controller=memory_controller)

    # Create an AssistantAgent, and attach teachability as its memory
    assistant_agent = AssistantAgent(
        name="teachable_agent",
        system_message = "You are a helpful AI assistant, with the special ability to remember user teachings from prior conversations.",
        model_client=client,
        memory=[teachability],
    )

    # Enter a loop to chat with the teachable agent
    print("Now chatting with a teachable agent. Please enter your first message. Type 'exit' or 'quit' to quit.")
    while True:
        user_input = input("\nYou: ")
        if user_input.lower() in ["exit", "quit"]:
            break
        await Console(assistant_agent.run_stream(task=user_input))

if __name__ == "__main__":
    import asyncio
    asyncio.run(main())

Head over to its README for details, and the samples for runnable examples.

New Sample: Gitty (Experimental)

Gitty is an experimental application built to help easing the burden on open-source project maintainers. Currently, it can generate auto reply to issues.

To use:

gitty --repo microsoft/autogen issue 5212

Head over to Gitty to see details.

Improved Tracing and Logging

In this version, we made a number of improvements on tracing and logging.

  • add LLMStreamStartEvent and LLMStreamEndEvent by @EItanya in #5890
  • Allow for tracing via context provider by @EItanya in #5889
  • Fix span structure for tracing by @ekzhu in #5853
  • Add ToolCallEvent and log it from all builtin tools by @ekzhu in #5859

Powershell Support for LocalCommandLineCodeExecutor

  • feat: update local code executor to support powershell by @lspinheiro in #5884

Website Accessibility Improvements

@peterychang has made huge improvements to the accessibility of our documentation website. Thank you @peterychang!

Bug Fixes

  • fix: save_state should not require the team to be stopped. by @ekzhu in #5885
  • fix: remove max_tokens from az ai client create call when stream=True by @ekzhu in #5860
  • fix: add plugin to kernel by @lspinheiro in #5830
  • fix: warn when using reflection on tool use with Claude models by @ekzhu in #5829

Other Python Related Changes

  • doc: update termination tutorial to include FunctionCallTermination condition and fix formatting by @ekzhu in #5813
  • docs: Add note recommending PythonCodeExecutionTool as an alternative to CodeExecutorAgent by @ekzhu in #5809
  • Update quickstart.ipynb by @taswar in #5815
  • Fix warning in selector gorup chat guide by @ekzhu in #5849
  • Support for external agent runtime in AgentChat by @ekzhu in #5843
  • update ollama usage docs by @ekzhu in #5854
  • Update markitdown requirements to >= 0.0.1, while still in the 0.0.x range by @afourney in #5864
  • Add client close by @afourney in #5871
  • Update README to clarify Web Browsing Agent Team usage, and use animated Chromium browser by @ekzhu in #5861
  • Add author name before their message in Chainlit team sample by @DavidYu00 in #5878
  • Bump axios from 1.7.9 to 1.8.2 in /python/packages/autogen-studio/frontend by @dependabot in #5874
  • Add an optional base path to FileSurfer by @husseinmozannar in #5886
  • feat: Pause and Resume for AgentChat Teams and Agents by @ekzhu in #5887
  • update version to v0.4.9 by @ekzhu in #5903

New Contributors

**Full Chang...

Read more

Contributors

ekzhu, taswar, and 11 other contributors
Loading
0