AutoGen

AutoGen is an open-source Python framework from Microsoft Research for building multi-agent AI applications. Multi-agent means multiple AI agents, each with a defined role, working together through conversation to complete tasks that would be too complex or too slow for a single model to handle alone. AutoGen gives developers the infrastructure to define those agents, connect them to tools and LLMs, and control how they pass messages to each other. It is one of the foundational frameworks that established how the industry thinks about orchestrating AI agents.

How AutoGen works

An agent in AutoGen is a software component that can send and receive messages, use an LLM to decide what to do next, and take actions such as running code or calling an API. Agents are the main building block. Here is how the system is structured:

• AssistantAgent: The reasoning unit. It takes a task, uses an LLM (such as GPT-4 or Claude) to think through it, and replies with a plan, code, or answer. It does not execute code itself.
• UserProxyAgent: The execution unit. It receives messages from the AssistantAgent, runs any code contained in them, and returns the output. It can also route the task back for correction if the code fails.
• AgentChat: The high-level API you use to define agents, set their roles, and start conversations between them. This is the entry point for most developers.
• Core: The lower-level, event-driven runtime under AgentChat. It handles asynchronous messaging between agents and supports distributed, cross-language deployments.
• Extensions: Optional connectors that let agents work with external systems, including MCP servers, Docker for sandboxed code execution, and Azure OpenAI.
• AutoGen Studio: A no-code web interface for prototyping multi-agent workflows without writing Python. Useful for testing team structures before coding them.

When you run a workflow, agents exchange messages in a loop. The AssistantAgent proposes a solution, the UserProxyAgent executes it, and the result comes back. This continues until the task is done or a stopping condition is met.

What you can build with AutoGen

• Automated code generation pipeline: Set up one agent to write Python functions and another to execute and test them, so the system iterates on bugs automatically without human intervention.
• Multi-agent RAG system: Assign one agent to retrieve documents from a knowledge base, another to synthesise the information, and a third to generate a final report, splitting a complex research task across specialised roles.
• Supply chain optimisation tool: Use a coding agent to write and run optimisation queries against operational data, with a safeguard agent reviewing outputs before they are surfaced to the user.
• Automated subtitle generator: Combine agents for transcription, translation, and timestamp formatting so a single prompt turns a video into a captioned, multilingual output.
• Human-in-the-loop approval workflow: Build a pipeline where agents complete drafts and proposals autonomously, but pause and wait for human review at defined checkpoints before proceeding.
• Travel planning assistant: Route user preferences to specialised agents that independently check weather, suggest destinations, and compile itineraries, then merge their outputs into a single plan.