AutoGen

AutoGen is an open-source Python (and .NET) framework from Microsoft Research for building applications where multiple AI agents collaborate to solve tasks. Instead of a single model handling everything, AutoGen lets developers compose specialized agents — an AssistantAgent that proposes solutions, a UserProxyAgent that represents the human or automates approvals, and optional tool or code-executor agents — that converse with one another, exchange code, critique proposals, and iterate until a task is complete. This conversation-driven paradigm makes agent reasoning transparent and easier to debug than opaque single-model chains.

The framework is organized in layers: a low-level Core API for event-driven, distributed agent runtimes; a higher-level AgentChat API for rapid prototyping of common multi-agent patterns like two-agent chat and group chats; and an Extensions API that plugs in LLM providers (OpenAI, Azure OpenAI, and other OpenAI-compatible endpoints), Model Context Protocol servers, and code execution sandboxes such as Docker. AutoGen Studio complements the code-first frameworks with a no-code, browser-based UI for assembling and testing agent teams visually, while AutoGen Bench provides a benchmarking suite for evaluating agent performance.

AutoGen is best suited for developers, researchers, and teams prototyping multi-agent workflows, experimenting with agent collaboration patterns, or teaching or studying how conversational multi-agent systems behave. Because it is one of the most cited open-source agent frameworks in academic literature, it is also a common baseline in multi-agent research papers.

Important context for 2026: Microsoft has placed AutoGen into maintenance mode. It receives only bug fixes and security patches and is now community-managed rather than actively developed by Microsoft. Microsoft's recommended path for new, production-oriented projects is Microsoft Agent Framework, which merges AutoGen's agent abstractions with Semantic Kernel's enterprise features. A community-led fork called AG2 also continues active development under Apache 2.0 for teams who want the classic AutoGen patterns with ongoing feature work. Existing AutoGen users can continue using the stable v0.4+ architecture, but should evaluate a migration path before starting new production builds.

AutoGen itself is completely free and open source under the MIT License (code) and CC-BY-4.0 (documentation). There are no paid tiers, seats, or usage fees for the framework. Costs come entirely from the LLM API calls agents make — for example OpenAI, Anthropic, or Azure OpenAI usage-based pricing — and any compute or Azure infrastructure used to host agent runtimes or Docker-based code execution. AutoGen Studio and AutoGen Bench are also free and installed via pip alongside the core packages.