AUTOGEN

AutoGen Microsoft multi-agent conversational framework.

Definition

AutoGen v0.2 (current stable 2024) key concepts: (1) ConversableAgent: base class agent that can send/receive messages, configurable LLM (OpenAI GPT-4, Claude, Llama via Ollama, etc.), system message, max turns. (2) AssistantAgent: pre-built agent acts as AI assistant, can suggest code execute, default GPT-4. (3) UserProxyAgent: pre-built agent acts as user proxy, can execute code + tools + invoke functions + ask human via human_input_mode. (4) GroupChat + GroupChatManager: group chat coordination multiple agents, manager selects next speaker based selection logic (round_robin, auto LLM-decides, custom function). (5) Code execution: agents can write + execute Python code locally Docker isolated environments, agents iterate fix errors. (6) Functions + Tools: register_function decorator expose Python functions to agents. (7) Caching: LLM response caching DiskCache + Redis options. AutoGen v0.4 (October 2024 major rewrite): refactored modular layered architecture (autogen-core foundational + autogen-agentchat conversational + autogen-ext extensions), production-grade improvements, async-first design, distributed multi-agent across processes/machines. Customers: Microsoft internal teams + research, ~10000+ GitHub Discussions + community contributions.

Origin

AutoGen announced August 2023 by Microsoft Research + Penn State University + University of Washington collaboration (Chi Wang + others, paper FLAML team) ; v0.1 OSS release ; v0.2 stable 2024 ; v0.4 major rewrite October 2024.

Example in context

Software development AutoGen group: agents = SeniorEngineer (Assistant GPT-4) + Reviewer (Assistant Claude) + UserProxy (executes code) ; user 'implement REST API for user management', SeniorEngineer writes Python FastAPI code, UserProxy executes + reports errors, Reviewer suggests improvements + security concerns, iterative until working code + Reviewer approval.

Related terms

• CrewAI — alternative multi-agent.