AutoGen is a multi-agent AI application framework for building conversational agent systems. It provides a comprehensive set of abstractions for agent communication, tool use, code execution, memory, and orchestration across both Python and .NET ecosystems. The framework includes a low-level runtime (autogen-core) with message passing, topic-based subscriptions, and pluggable runtimes (in-process, gRPC, Orleans), and a high-level agent chat library (autogen-agentchat) with opinionated agent types (AssistantAgent, UserProxyAgent, CodeExecutorAgent), team orchestration patterns (RoundRobin, Selector, Swarm, MagenticOne, Graph), and composable termination conditions. Extensions (autogen-ext) provide integrations with major LLM providers (OpenAI, Anthropic, Gemini, Ollama, llama.cpp, etc.), code executors (local, Docker, Jupyter, Azure), memory backends (ChromaDB, Redis, mem0, canvas), and tools (MCP, LangChain, HTTP, GraphRAG). The repository also includes AutoGen Studio, a full-stack web application (FastAPI + React) for visually building and managing agent teams, and agbench, a benchmarking tool for evaluating agent performance on tasks like HumanEval and GAIA. The architecture is modular: core runtime handles agent lifecycle and message routing; agent chat provides higher-level conversational patterns; extensions offer pluggable implementations. Key entry points include the SingleThreadedAgentRuntime and GrpcAgentRuntime for running agents, the AssistantAgent and Team classes for building chat applications, and the autogenstudio CLI for launching the web UI. Data flows involve agents sending messages via topics or direct addressing, with middleware and intervention handlers for customization. External dependencies include LLM APIs, Docker, Redis, ChromaDB, and various cloud services. The repository is archived and no longer maintained, with migration guidance to Microsoft Agent Framework (MAF). Despite this, the codebase demonstrates high craftsmanship with clean abstractions, extensive testing, and reusable patterns. It is a valuable reference for multi-agent system design.