What Is AutoGen and Microsoft Agent Framework

The Origins of AutoGen

AutoGen began as a research project at Microsoft Research, first published in a paper in August 2023. The researchers observed that many AI tasks were too complex for a single language model to handle reliably in one pass. Tasks like data analysis, software development, and research synthesis naturally involve multiple perspectives, iterative refinement, and different types of expertise. AutoGen's solution was to model these workflows as conversations between multiple AI agents, each with its own role and capabilities.

The project was released as open-source software under the MIT license and rapidly attracted attention from the AI development community. By early 2026, AutoGen had accumulated over 54,000 GitHub stars, making it one of the most popular multi-agent frameworks available. Its success demonstrated that the conversational multi-agent approach resonated with developers building real applications.

The name "AutoGen" reflects the framework's core philosophy: automatically generating solutions through agent collaboration. Rather than requiring developers to manually code every step of a complex workflow, AutoGen enables agents to figure out the steps through conversation, adapting their approach based on intermediate results and feedback.

Core Concepts and How It Works

At the heart of AutoGen are three types of agents that work together. The AssistantAgent is an AI-powered agent that uses a language model to understand tasks, generate plans, write code, and produce analysis. It receives instructions through its system message, which defines its personality, expertise, and constraints. Developers create multiple AssistantAgents with different system messages to simulate a team of specialists.

The UserProxyAgent acts as the bridge between AI agents and the human world. It can relay human input into the conversation, execute code written by other agents in sandboxed environments, invoke external tools, and apply approval policies that determine whether actions require human confirmation. In fully autonomous mode, the UserProxyAgent runs without human intervention, while in supervised mode, it pauses for approval before executing potentially sensitive operations.

The GroupChatManager coordinates conversations between three or more agents. It maintains the shared conversation history, determines which agent should speak next (either through round-robin, LLM-based selection, or custom logic), and enforces termination conditions that end the conversation when the task is complete or resource limits are reached.

A typical AutoGen workflow starts when a user provides a task to the system. The UserProxyAgent passes this task to an AssistantAgent, which analyzes the request, formulates a plan, and may write Python code to accomplish part of the task. The code goes back to the UserProxyAgent for execution in a sandboxed environment. The execution output (whether successful results or error messages) returns to the AssistantAgent, which interprets the results and either declares the task complete or generates improved code for another attempt. This iterative cycle continues until the task succeeds or a maximum number of conversation turns is reached.

What Makes AutoGen Different

AutoGen's conversational approach distinguishes it from other multi-agent frameworks in several important ways. First, the conversation history serves as the shared state for all agents. There is no separate database, state machine, or workflow engine tracking progress. Every agent can read the full transcript of the conversation and use it to make informed decisions about what to do next. This simplicity makes AutoGen easier to understand and debug than frameworks that require explicit state management.

Second, AutoGen treats code execution as a first-class citizen. When an AI agent generates code, the framework can automatically extract it, run it in a sandboxed environment, and feed the results back into the conversation. This creates a powerful iterative debugging loop where agents write code, observe errors, and fix problems across multiple turns, much like a human developer working in a REPL environment.

Third, AutoGen is model-agnostic. While many agent frameworks are tightly coupled to a specific LLM provider, AutoGen works with OpenAI GPT models, Azure OpenAI, Anthropic Claude, Google Gemini, and open-source models through compatible APIs. Different agents in the same system can even use different models, allowing developers to optimize cost and performance by matching model capabilities to task requirements.

Fourth, the framework supports flexible conversation topologies. Beyond simple two-agent dialogues, AutoGen enables group chats with dynamic speaker selection, nested conversations where agent pairs handle sub-tasks, and hierarchical orchestration patterns where manager agents coordinate teams of specialists. These patterns can be combined to model complex organizational structures and decision-making processes.

The Transition to Microsoft Agent Framework

In October 2025, Microsoft announced a significant strategic shift. The AutoGen team and the Semantic Kernel team would merge to create the Microsoft Agent Framework, a unified SDK for building agent-based AI applications. AutoGen entered maintenance mode, meaning it would continue to receive bug fixes and security patches but no major new features.

The Microsoft Agent Framework (MAF) reached version 1.0 GA in April 2026. It preserves AutoGen's conversational multi-agent patterns while adding enterprise-grade features from Semantic Kernel: a plugin system for tool integration, persistent memory with vector stores, session-based state management, OpenTelemetry-based observability, and cross-runtime interoperability through the A2A and MCP protocols.

The framework supports both Python and .NET, with agents written in either language able to communicate through the Agent-to-Agent protocol. This cross-runtime capability is particularly valuable for enterprise organizations with polyglot development teams and existing investments in the .NET ecosystem.

For developers, the transition means that new projects should start with the Microsoft Agent Framework rather than AutoGen. The core concepts are the same, and Microsoft provides a migration guide for existing AutoGen projects. The migration is primarily a refactoring exercise involving updated import statements and configuration, not a complete rewrite of agent logic.

Who Should Use AutoGen

AutoGen and the Microsoft Agent Framework are well-suited for developers building applications that involve multi-step reasoning, code generation and execution, data analysis, or coordination between specialized AI capabilities. Common use cases include automated research assistants that gather and synthesize information, coding agents that write, test, and debug software, data analysis pipelines that clean, transform, and visualize datasets, and customer service systems where different agents handle different aspects of customer interactions.

The framework is particularly valuable for teams already invested in the Microsoft ecosystem, including Azure, .NET, and Microsoft 365. The deep integration with Azure AI services, the availability of both Python and .NET SDKs, and the forthcoming Agent 365 product make it a natural choice for enterprise organizations with existing Microsoft infrastructure.

Teams that prioritize model flexibility will also benefit from AutoGen's provider-agnostic design. Unlike frameworks tied to a single LLM vendor, AutoGen and MAF support switching between model providers without changing agent logic, reducing vendor lock-in risk and enabling cost optimization through provider competition.

Developers who need strong code execution capabilities will find AutoGen's sandboxed execution environment and iterative debugging loop more mature than alternatives. The Docker-based executor provides production-grade isolation, and the conversation-driven error correction pattern handles common coding failures gracefully without manual intervention.