Open-Source AgentOS for AI Agents — a framework and OS-level platform for building, orchestrating, and evolving multi-agent collaborative systems, formerly known as Microsoft AutoGen.
Core Architecture#
AG2 is built around ConversableAgent as the unified base class, with specialized types like AssistantAgent (LLM-driven) and UserProxyAgent (human proxy). The orchestration layer resides in autogen.agentchat, with AutoPattern providing automatic agent selection. The current stable framework will be gradually deprecated in favor of autogen.beta for the upcoming v1.0. Data validation is based on Pydantic 2, and the build system uses pyproject.toml + setup.jinja templates.
Multi-Agent Orchestration#
- Orchestration patterns: group chat, nested chat, sequential chat, swarm, etc.; Pattern Cookbook provides 9 group orchestration patterns
- AutoPattern: automatic agent selection
- A2A Protocol: supports Agent-to-Agent communication protocol
Runtime Capabilities#
- Code execution: built-in, with Docker isolation support
- RAG: built-in support, compatible with Qdrant, PGVector, MongoDB, Couchbase, FalkorDB, etc.
- Tool/function calling:
register_functionAPI for custom Python functions, MCP (Model Context Protocol) compatible - Browser integration: browser-use, crawl4ai, websurfer
- Structured output: supports structured format agent output
- Human-in-the-Loop: via
UserProxyAgentwithNEVER/ALWAYS/TERMINATEinput modes,is_termination_msgcallback for termination control
LLM Backend Support#
OpenAI, Anthropic, Google Gemini, Bedrock, Cerebras, Cohere, DeepSeek, Groq, Mistral, Ollama, Together, and more.
Cross-Framework Interoperability & Communication#
Interoperates with LangChain, CrewAI, Pydantic AI, Google ADK, and other frameworks. Built-in integrations for Discord, Slack, and Telegram.
Typical Use Cases#
- Multi-agent coding collaboration (Coder + Reviewer pattern)
- Education and content creation with multi-role collaboration
- Math problem solving (MathChat pattern, paper-supported)
- Enterprise workflow automation (claimed up to 70% workflow speed improvement and 5x productivity)
- Cross-framework agent team orchestration
- Tool-augmented intelligent assistants
Installation & Quick Start#
Requirements: Python >= 3.10, < 3.14
pip install ag2[openai]
# or using alias
pip install autogen[openai]
Optional extras: a2a, anthropic, browser-use, mcp, rag, redis, neo4j, interop, etc.
from autogen import AssistantAgent, UserProxyAgent, LLMConfig
llm_config = LLMConfig.from_json(path="OAI_CONFIG_LIST")
assistant = AssistantAgent("assistant", llm_config=llm_config)
user_proxy = UserProxyAgent("user_proxy", code_execution_config={"work_dir": "coding", "use_docker": False})
user_proxy.run(assistant, message="Summarize the main differences between Python lists and tuples.").process()
Note: PyPI package name is ag2, but Python import name remains autogen.
Unconfirmed Information#
- v1.0 release timeline not explicitly stated
- AG2AI organization governance details not publicly available
- Enterprise products (Orchestrator / Studio / Applications) have no public pricing
- PyPI lists version 0.12.0 as released "Apr 17, 2026" — this date may be a system display anomaly
autogen.betamodule API changes and migration path require consulting the docs site