AutoGen

AutoGen is an open-source framework developed by Microsoft Research that allows developers to build applications using multiple large language model (LLM)-powered agents that communicate with each other through structured conversation to solve complex tasks. First published in 2023 alongside an academic paper demonstrating its capabilities across domains including mathematics, coding, supply-chain optimisation, and question answering, AutoGen became one of the most widely adopted multi-agent frameworks in the research and developer community before transitioning to maintenance mode as Microsoft introduced its successor, the Microsoft Agent Framework.

Core Concepts

AutoGen is built around the concept of conversable agents: autonomous units that can send and receive messages, generate responses using an LLM, execute code, or call tools, and maintain conversation history. The framework abstracts the complexity of multi-agent coordination into a conversation-centric paradigm — rather than defining explicit control flow as code, developers define agents and their response policies, and the conversation among agents drives the task to completion organically.

Two fundamental agent types underpin AutoGen: the AssistantAgent, which wraps an LLM and generates responses based on conversation history and its system prompt, and the UserProxyAgent, which can represent a human participant, execute code on the human's behalf, or provide scripted responses. The human-in-the-loop model allows developers to configure how much autonomy agents have — from fully automated execution to requiring human approval at each step.

Group Chat and Orchestration

Beyond simple two-agent conversations, AutoGen supports group chat scenarios in which multiple agents participate in a shared conversation thread. A GroupChatManager agent moderates turn-taking, deciding which agent should respond next based on the conversation state. This enables patterns such as a manager agent directing specialist agents, a panel of diverse agents debating alternative approaches, or a team of specialist agents each contributing domain expertise to produce a synthesised result.

The framework supports code execution through a sandboxed environment, allowing agents to write Python code in response to a task, execute it, observe the output, and revise their approach based on the result — a loop particularly effective for data analysis, mathematical reasoning, and software engineering tasks.

Applications

AutoGen was demonstrated to be effective across a broad range of applications in its introductory research paper. In mathematics, agent conversations guided step-by-step reasoning through competition-level problems. In coding, teams comprising a coder, a reviewer, and an executor collaborated to produce working software from natural-language specifications. In question answering, retrieval agents fetched documents and reasoning agents synthesised answers. In supply-chain optimisation, agents with access to simulation tools ran experiments and refined strategies iteratively.

These demonstrations established AutoGen as a practical framework for tasks that benefit from iterative refinement, error recovery through code re-execution, and specialisation — properties that naturally emerge from multi-agent conversation when each agent has a well-defined role and access to appropriate tools.

Relationship to Microsoft Agent Framework

Microsoft announced in 2025 that AutoGen would transition to maintenance mode, continuing to receive bug fixes and security updates but no new features. The Microsoft Agent Framework is a multi-language SDK supporting Python, C#, and TypeScript that extends and generalises the ideas pioneered in AutoGen with improved scalability, observability, and enterprise integration features. A migration guide published on Microsoft Learn assists developers in porting AutoGen applications to the new framework.

Despite this transition, AutoGen's GitHub repository remains among the most-starred AI agent repositories, and a large body of tutorials, blog posts, academic papers, and production deployments continue to reference it. Its influence on subsequent multi-agent frameworks — including CrewAI and LangGraph — is well documented.

Design Comparisons

AutoGen's conversation-centric design contrasts with CrewAI's role-and-crew metaphor and LangGraph's graph-based workflow representation. Each approach reflects different priorities: AutoGen emphasises flexibility and emergent behaviour through unconstrained agent dialogue; CrewAI emphasises role clarity and accessibility for business workflow designers; LangGraph emphasises deterministic control flow for production reliability. Developers selecting between these frameworks typically consider the degree of workflow predictability required, the importance of human oversight at specific steps, and the availability of engineering resources to manage the respective abstractions.

References

1. Wu, Q. et al. (2023). AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation Framework. Microsoft Research.
2. Microsoft. (2025). AutoGen to Microsoft Agent Framework Migration Guide. Microsoft Learn.
3. GitHub. (2024). microsoft/autogen: A programming framework for agentic AI. github.com/microsoft/autogen.
4. Tribe AI. (2024). Microsoft AutoGen: Orchestrating Multi-Agent LLM Systems. tribe.ai.
5. Microsoft Research. (2023). AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. Microsoft Research Blog.