AutoGen

AutoGen is Microsoft Research’s open-source framework for building multi-agent AI applications, where multiple AI agents hold conversations with each other and with humans to complete tasks. Instead of writing one giant prompt for a single model, you define specialist agents (a coder, a reviewer, a planner) and let them collaborate through structured dialogue until the task is finished. Microsoft has since moved AutoGen into maintenance mode in favor of Microsoft Agent Framework, but its conversational multi-agent pattern remains one of the clearest ways for AI developers to understand how agents can reason, critique each other’s work, and hand off tasks in a production system.

How AutoGen works

AutoGen treats agent collaboration as a conversation. Each agent is an object with a role, a set of tools, and a connection to an LLM (large language model). Agents send messages to each other, and the framework manages the back-and-forth until the group reaches a solution or hits a stopping condition.

• Conversable agents: Every agent in AutoGen inherits from a base class that can send and receive messages, call tools, and optionally hand off to a human. This makes agents composable, since any agent can talk to any other agent through the same interface.
• AssistantAgent and UserProxyAgent: The most common pairing is an AssistantAgent, which calls an LLM to generate responses, and a UserProxyAgent, which can execute code, run tools, or ask a human for input. Together they form a loop where the assistant proposes and the proxy verifies or acts.
• Group chat: For tasks that need more than two agents, AutoGen’s group chat manager routes messages between multiple agents in a shared conversation, selecting who speaks next based on the current state of the discussion.
• Tool and MCP support: Agents can call external tools directly or connect to Model Context Protocol (MCP) servers, an open standard for linking AI models to external data sources and services, to pull in live data or take real-world actions.
• Layered architecture: The Core API handles low-level message passing and distributed runtime. The AgentChat API sits on top with a simpler interface for common patterns like two-agent chat or group chat. The Extensions API adds first and third-party integrations, such as OpenAI and Azure OpenAI model clients.

What you can build with AutoGen

• Code generation and debugging pipeline: Pair a coder agent that writes code with an executor agent that runs it in a sandbox, reports errors back, and lets the coder revise its work until the tests pass.
• Multi-agent research assistant: Assign one agent to search the web, another to fact-check claims, and a third to draft a summary, then have them critique each other’s output before the final report ships.
• Customer support triage bot: Build a group chat where a classifier agent routes tickets, a specialist agent drafts a resolution, and a human-in-the-loop proxy agent approves or edits the response before it goes out.
• Educational tutoring app: Have a teaching agent generate explanations and a student agent ask clarifying questions, surfacing gaps in the explanation before it ever reaches a real learner.
• Data analysis workflow: Chain an agent that writes analysis code, an agent that executes it against a dataset, and an agent that interprets the results into a plain-English summary for a non-technical stakeholder.