Autogen

AutoGen is Microsoft Research's multi-agent framework with a strong research pedigree and broad community adoption. Its conversational agent model (agents exchange messages to collaboratively solve tasks) supports complex topologies: two-agent conversations, group chats with a manager, human-in-the-loop interruption, and code execution sandboxes. AutoGen 0.4 (AgentChat API) is the current recommended interface — cleaner than 0.2. Good for teams that need research-grade flexibility with practical deployment paths. Confidence is docs-derived.

Python SDK: pip install autogen-agentchat (v0.4). AssistantAgent, UserProxyAgent, GroupChat, GroupChatManager are the core primitives. Agent conversations via agent.initiate_chat(recipient, message). Code execution: UserProxyAgent with code_execution_config runs LLM-generated code in a sandbox. AutoGen Studio: no-code UI for agent design. AutoGen Core for lower-level event-driven agent architecture.

No AutoGen-specific API key. LLM provider configured via OAI_CONFIG_LIST env variable or config dict. Code execution sandbox: Docker or local subprocess (local has security implications). Open-source MIT license. No cloud auth required for self-run usage.

Code execution sandboxing is the key reliability concern — local subprocess execution without Docker exposes host system to LLM-generated code. Docker sandbox recommended for production. Agent conversation loops can run indefinitely — configure max_turns or human_input_mode="TERMINATE" guard. Python exceptions surface on LLM API failures. Active maintainership from Microsoft Research.

microsoft.github.io/autogen covers quickstart, agent design patterns, code execution, group chat, and migration guides from v0.2 to v0.4. Getting started: pip install autogen-agentchat, configure LLM, first two-agent conversation in under 10 minutes. Active Discord community and GitHub. Extensive academic citations. Tutorials on Microsoft Learn.