microsoft/agent-framework
A framework for building, orchestrating and deploying AI agents and multi-agent workflows with support for Python and .NET.
What it solves
Microsoft Agent Framework (MAF) 提供一致且具生產等級的基礎，讓開發者能建置與營運 AI 代理與多代理工作流程。它彌補了簡易原型與正式系統之間的差距，提供編排、耐久性與可觀測性等在基本聊天迴圈中常缺乏的工具。
How it works
MAF 是一個支援 Python 與 .NET 的多語言框架。開發者可以使用圖形化模式（例如順序、並行與群組協作）來定義代理並編排它們，以處理複雜任務。框架支援多種 LLM 供應商（包括 Azure OpenAI 與 OpenAI），並內建中介層系統供自訂請求/回應處理。它亦提供透過 YAML 的宣告式代理定義，以及專用的 DevUI 供測試與除錯。
Who it’s for
此框架設計給需要從無狀態聊天迴圈升級，建構耐久、可觀測且受治理的代理系統，並能部署至本機或雲端環境的開發團隊。
Highlights
- Multi-language support: Consistent APIs across Python and C#/.NET.
- Graph-based orchestration: Supports complex workflows including handoffs, concurrent execution, and human-in-the-loop control.
- Production readiness: Built-in OpenTelemetry integration for monitoring and features like checkpointing and time-travel.
- Flexible deployment: Supports various hosting patterns, including Foundry-hosted agents, Azure Functions, and Durable Task hosting.
- Extensible architecture: Includes a middleware system and support for multiple LLM providers to prevent vendor lock-in.