AutoGen

AutoGen is an open-source multi-agent framework originally developed by Microsoft Research for building applications where multiple AI agents converse, collaborate, and execute code together. In 2025, AutoGen underwent a significant evolution — rebranding as the Microsoft Agent Framework (also known as AG2) — with version 0.4 introducing asynchronous messaging, modular extensibility, and distributed scalability.

The framework has over 20,000 active builders and remains one of the most widely used platforms for multi-agent AI development.

History and Rebrand

AutoGen was initially released by Microsoft Research as a framework for multi-agent conversations. The original design featured two layers: a low-level autogen-core with RoutedAgent and message handlers, and a high-level Team API with patterns like RoundRobinGroupChat. Community feedback indicated that the low-level API was overly complex while the high-level API was too restrictive.

In January 2025, AutoGen v0.4 was released, and the project was rebranded to AG2 and later integrated into the Microsoft Agent Framework. This migration unified the architecture around Workflow and Magentic builders, simplifying development while retaining power.

Key Features

Multi-Agent Conversations: Support for two-agent chats, group chats with dynamic speaker selection, sequential and nested conversations, and sophisticated message routing.

Code Execution: Integrated code execution capabilities through the extensions API, allowing agents to write, run, and debug code autonomously.

Group Chat: Dynamic coordination with group chat managers, termination conditions, and support for specialized roles (critic, executor, planner).

Human-in-the-Loop: Built-in support for stall handling, plan review, and approval requests, enabling human oversight at critical decision points.

Observability: OpenTelemetry integration for tracing, metrics, and debugging across agent workflows.

AutoGen Studio / AG2 Studio: A UI-based interface for visually building and testing multi-agent workflows without extensive code.

Architecture

The Microsoft Agent Framework uses a layered architecture:

Core API — Event-driven message passing with RoutedAgent, supporting both local and distributed runtimes
AgentChat API — High-level multi-agent patterns built on Core (teams, chats, group coordination)
Extensions API — LLM clients (OpenAI, Azure), code execution, and community tools
Workflow Builders — MagenticBuilder for orchestrated multi-agent teams and WorkflowBuilder for edge-based execution graphs

Code Example

from autogen import AssistantAgent, UserProxyAgent
 
config_list = [{"model": "gpt-4o", "api_key": "your-key"}]
 
assistant = AssistantAgent(
    name="assistant",
    llm_config={"config_list": config_list}
)
 
user_proxy = UserProxyAgent(
    name="user_proxy",
    human_input_mode="TERMINATE",
    code_execution_config={"work_dir": "coding", "use_docker": False}
)
 
user_proxy.initiate_chat(
    assistant,
    message="Write a Python function to calculate fibonacci numbers and test it."
)

Integration with Semantic Kernel

The Microsoft Agent Framework converges with Semantic Kernel through three integration paths: AutoGen agents can leverage Semantic Kernel capabilities, AutoGen agents can be hosted within the SK ecosystem, and harmonized runtimes enable multi-agent interoperability. This unified approach supports deployment through Azure AI Foundry Agent Service.

Table of Contents