微软近日发布了Microsoft Agent Framework（公测版），这是一个开源的SDK和运行时，整合了AutoGen和Semantic Kernel的核心理念。该框架旨在帮助团队构建、部署和监控生产级的AI智能体及多智能体工作流。支持Python和.NET，并与Azure AI Foundry的Agent Service集成，提供可扩展性和运营支持。Agent Framework统一了智能体运行时和API接口，融合了AutoGen的多智能体抽象和Semantic Kernel的企业级功能，如状态管理、类型安全和遥测。它支持首级的编排模式，包括智能体编排和工作流编排，并提供Pro-code开发能力和平台互操作性，允许开发者灵活选择模型提供商，降低供应商锁定。

🚀 **统一的AI智能体开发体验**: Microsoft Agent Framework整合了AutoGen在多智能体模式和运行时方面的优势，以及Semantic Kernel在企业控制、状态管理和插件集成方面的能力，为开发者提供了一个统一的平台来构建、部署和监控生产级的AI智能体和多智能体工作流。

💡 **灵活的智能体与工作流编排**: 该框架支持两种首级的编排模式：智能体编排（LLM驱动的决策制定）和工作流编排（确定性的、基于业务逻辑的多智能体流程）。这使得开发者能够构建混合系统，将创意规划与可靠的任务交接和约束相结合。

🌐 **开放、多语言的SDK与互操作性**: Microsoft Agent Framework提供基于MIT许可的Python和.NET的开源SDK，方便开发者集成。其AIAgent接口设计支持与Azure AI Foundry Agents、OpenAI Assistants和Copilot Studio互操作，并允许轻松更换聊天模型提供商，有效降低了应用层的供应商锁定。

🔧 **面向生产环境的强大支持**: 结合Azure AI Foundry的Agent Service，该框架提供了托管运行时，能够处理模型链接、工具集成、线程状态管理、内容安全和身份验证，并提供可观察性。这特别适用于复杂的、需要Pro-code开发的企业级场景，优化了AI经济中的关键因素，如吞吐量、延迟和故障恢复。

🔄 **可复现的状态管理与模型灵活性**: Agent Framework以“线程”作为状态管理的基本单位，支持可复现的运行、重试和审计。同时，其模型/提供商的灵活性允许开发者在不重写编排逻辑的情况下，针对不同任务选择Azure OpenAI、OpenAI、本地运行时（如Ollama）等多种模型，便于成本和性能的优化。

Microsoft released the Microsoft Agent Framework (public preview), an open-source SDK and runtime that unifies core ideas from AutoGen (agent runtime and multi-agent patterns) with Semantic Kernel (enterprise controls, state, plugins) to help teams build, deploy, and observe production-grade AI agents and multi-agent workflows. The framework is available for Python and .NET and integrates directly with Azure AI Foundry’s Agent Service for scaling and operations.

What exactly is Microsoft shipping?

A consolidated agent runtime and API surface. The Agent Framework carries forward AutoGen’s single- and multi-agent abstractions while adding Semantic Kernel’s enterprise features: thread-based state management, type safety, filters, telemetry, and broad model/embedding support. Microsoft positions it as the successor built by the same teams, rather than a replacement that abandons either project.First-class orchestration modes. It supports agent orchestration (LLM-driven decision-making) and workflow orchestration (deterministic, business-logic multi-agent flows), enabling hybrid systems where creative planning coexists with reliable handoffs and constraints.Pro-code and platform interoperability. The base AIAgent interface is designed to swap chat model providers and to interoperate with Azure AI Foundry Agents, OpenAI Assistants, and Copilot Studio, reducing vendor lock-in at the application layer. Open-source, multi-language SDKs under MIT license. The GitHub repo publishes Python and .NET packages with examples and CI/CD-friendly scaffolding. AutoGen remains maintained (bug fixes, security patches) with guidance to consider Agent Framework for new builds.

Where it runs in production?

Azure AI Foundry’s Agent Service provides the managed runtime: it links models, tools, and frameworks; manages thread state; enforces content safety and identity; and wires in observability. It also supports multi-agent orchestration natively and distinguishes itself from Copilot Studio’s low-code approach by targeting complex, pro-code enterprise scenarios.

But how is it connected to ‘AI economics’?

Enterprise AI economics are dominated by token throughput, latency, failure recovery, and observability. Microsoft’s consolidation addresses those by (a) giving one runtime abstraction for agent collaboration and tool use, (b) attaching production controls—telemetry, filters, identity/networking, safety—to the same abstraction, and (c) deploying onto a managed service that handles scaling, policy, and diagnostics. This reduces the “glue code” that typically drives cost and brittleness in multi-agent systems and aligns with Azure AI Foundry’s model-catalog + toolchain approach.

Architectural notes and developer surface

Runtime & state: Agents coordinate via a runtime that handles lifecycles, identities, communication, and security boundaries—concepts inherited and formalized from AutoGen. Threads are the unit of state, enabling reproducible runs, retries, and audits.Functions & plugins: The framework leans on Semantic Kernel’s plugin architecture and function-calling to bind tools (code interpreters, custom functions) into agent policies with typed contracts. (Model/provider flexibility: The same agent interface can target Azure OpenAI, OpenAI, local runtimes (e.g., Ollama/Foundry Local), and GitHub Models, enabling cost/performance tuning per task without rewriting orchestration logic.

Enterprise context

Microsoft frames the release as part of a broader push toward interoperable, standard-friendly “agentic” systems across Azure AI Foundry—consistent with prior statements about multi-agent collaboration, memory, and structured retrieval. Expect tighter ties to Foundry observability and governance controls as these stabilize.

Our Comments

We like this direction because it collapses two divergent stacks—AutoGen’s multi-agent runtime and Semantic Kernel’s enterprise plumbing—into one API surface with a managed path to production. The thread-based state model and OpenTelemetry hooks address the usual blind spots in agentic systems (repro, latency tracing, failure triage), and Azure AI Foundry’s Agent Service takes on identity, content safety, and tool orchestration so teams can iterate on policies instead of glue code. The Python/.NET parity and provider flexibility (Azure OpenAI, OpenAI, GitHub Models, local runtimes) also make cost/perf tuning practical without rewriting orchestration.

The post Microsoft Releases ‘Microsoft Agent Framework’: An Open-Source SDK and Runtime that Simplifies the Orchestration of Multi-Agent Systems appeared first on MarkTechPost.