2025年5月25日

微软 Build 2025：迈向智能体网络时代

Microsoft Build 2025: Ushering in the Agentic Web

In mid‑May 2025, at the annual Microsoft Build conference in Seattle (May 19–22), Microsoft unveiled more than 50 new AI tools, marking a transformative shift toward what it calls the open agentic web. Under the leadership of CEO Satya Nadella and CTO Kevin Scott, Microsoft showcased an ecosystem of products—spanning Windows, Azure, GitHub, Microsoft 365, and more—designed to empower AI agents that act autonomously, collaborate with other agents, and integrate seamlessly into real-world workflows.

1. The Vision: From Passive Assistants to Autonomous Agents

The core message at Build 2025 was clear: Microsoft believes we’re entering an era dominated by agentic AI—intelligent systems that do more than respond; they proactively plan, decide, and execute. Kevin Scott noted that usage of AI agents has more than doubled in the past year, highlighting growing adoption of systems that handle delegated human tasks Business Insider.

Microsoft described the emerging open agentic web using an analogy to past shifts in computing—like the rise of graphical interfaces or the internet—but centered on natural‑language interaction. In this view, AI agents plugged into platforms become the new interface layer. A foundational aspect is interoperability, akin to how USB‑C unified power and data connections. To that end, Microsoft introduced two open standards: the Model Context Protocol (MCP) and NLWeb, each designed to unlock agent-to-agent communication and natural‑language interfaces for the web.

2. Open Standards: MCP and NLWeb

2.1 Model Context Protocol (MCP)

MCP was originally defined by Anthropic in late 2024 and rapidly adopted by industry leaders. It establishes a universal JSON‑RPC‑based interface for language models to access data sources, tools, and context—a true “USB‑C for AI.” In the past months, OpenAI, Google DeepMind, and others have also embraced MCP.

At Build, Microsoft announced native MCP support across its intelligent stack: this includes Azure AI Foundry, Copilot Studio, Dynamics 365, GitHub, Semantic Kernel, and Windows 11 itself. MCP enables applications—including system services, file access, and cloud-based datasets—to be accessed securely by AI agents.

Windows AI Foundry, in particular, will support MCP via a secure registry and server framework, allowing AI agents to perform authorized actions on local applications and services. This infrastructure includes consent prompts, principle-of-least-privilege controls, and private developer preview access to select partners .

A recent academic report also underscores the importance of securing MCP workflows. It highlights risks like malicious tool servers and prompt injection, and proposes monitoring frameworks to preserve safe interactions .

2.2 NLWeb

Ramanathan Guha, a veteran creator of RSS and Schema.org, introduced NLWeb as a protocol that lets any website embed conversational AI powered by its own content and choice of model. Instead of relying on external bots, a developer can launch a natural-language chatbot interface on their own site with minimal effort, using syndicated content and customizable vector search .

Demonstrations at Build featured platforms like Serious Eats, TripAdvisor, and Shopify deploying NLWeb to deliver tailored experiences. For example, a visitor could specify dietary preferences like “vegetarian recipes for Diwali,” and the system remembers that context. Another example showcased searching outdoor gear with awareness of weather or brand preferences, dynamically surfacing relevant results.

Microsoft positions NLWeb and MCP as complementary open standards: NLWeb enables conversational interfaces on the web, while MCP enables agentic control and interoperability behind the scenes.

3. GitHub Copilot: Autonomous Engineering Agent

At the heart of Microsoft’s AI-first developer pitch is the evolution of GitHub Copilot. No longer a passive code suggestion tool, Copilot is now being positioned as a fully autonomous engineering agent.

Key capabilities introduced include:

Agent Workflows: Copilot can triage issues, write code, refactor or fix bugs, generate unit tests, and create draft pull‑requests independently.
Azure SRE Agent: Within GitHub Copilot, a built-in Site Reliability Engineering agent monitors system health, addresses incidents automatically, and proactively manages deployment reliability.
Copilot Studio Integration: Developers can build multi‑agent pipelines in Copilot Studio, combining specialized agents that fetch data, generate user-facing outputs, or schedule tasks. This integrates seamlessly with Azure AI Foundry and the A2A protocol for agent-to-agent coordination Medium VentureBeat.

The result is a vision of Copilot not as a helper, but as a collaborative coding colleague—capable of managing entire workflows from issue to release with minimal human oversight.

4. Azure AI Foundry: The Agent Cloud

Microsoft’s centerpiece for enterprise AI is Azure AI Foundry, a comprehensive platform for building, deploying, and managing AI agents at scale.

4.1 Platform Highlights

1,900+ Models Available: The Foundry now supports more than 1,900 models—from Microsoft and third parties—including xAI’s Grok 3 and Grok 3 Mini. Integration with Elon Musk’s Grok was formally announced during Build, despite broader tensions, highlighting strategic cooperation in AI development .
Agent Service GA: The Azure AI Foundry Agent Service officially reached general availability. It provides a fully managed environment for creating enterprise-grade, multi-agent workflows—supporting protocols like Agent-to-Agent (A2A) and MCP, with debugging and trace capabilities .
Model Router & Leaderboard: A new feature automates selection of optimal models for tasks and tracks performance metrics — letting developers tune efficiency and cost across multiple models .
Code Interpreter Preview: Developers can build agents that generate and execute Python code for tasks like data analysis or visualization—ideal for operations like RAG (Retrieval Augmented Generation) workflows.
Agentic Retrieval in Azure AI Search: A preview feature where agents orchestrate queries using conversational context and planning—prioritizing domain knowledge and history.
Observability & Governance: Azure AI Foundry Observability is launching a dashboard that brings real-time metrics, tracing, cost analysis, and safety checks—allowing developers to monitor and control agent behavior end-to-end .

4.2 Local and Edge Deployment: Foundry Local

Microsoft also announced Azure AI Foundry Local, aimed at running AI agents directly on user devices. Available on Windows 11 and macOS, Foundry Local uses ONNX Runtime to deliver on-device inference—reducing latency, cost, and privacy exposure .

5. Windows AI Foundry & MCP Native Support

Microsoft introduced Windows AI Foundry as the desktop-facing extension of Azure AI Foundry. This brings agentic capability to PCs in several ways:

5.1 Core Offerings

Foundry Local Catalog: Developers can access, test, and run open-source models optimized across CPUs, GPUs, and NPUs on Windows via the Foundry Local CLI or built-in SDKs.
Native MCP Support: Windows now includes an MCP registry and server architecture allowing AI agents to interface with core OS features such as file systems, WSL, and app-specific actions. A private developer preview will open access to select partners.
Developer Tools and Security Enhancements: New additions include the VBS Enclave SDK, post‑quantum cryptography support, and App Actions—tools designed for injecting agent-friendly hooks and ensuring secure agent usage.
Open-Source WSL: As part of the push toward transparency and community-driven development, Microsoft open-sourced the Windows Subsystem for Linux (WSL), including UI and underlying components.
Updated Developer Utilities: Tools like Terminal, WinGet, PowerToys, and Microsoft Store workflows were refreshed to align with the AI-first direction for Windows developers .

5.2 Strategic Implications

This integration reframes Windows PCs as full-fledged agents environments—machines capable of running AI services locally, while interoperating with cloud capabilities. Microsoft portrays this as bridging the gap between on-device responsiveness, user privacy, and enterprise governance.

6. Copilot Studio: Low-Code Agent Builder

Copilot Studio, Microsoft’s visual agent orchestration tool, received substantial upgrades:

Multi-Agent Orchestration (Private Preview): Users can now define pipelines of cooperating agents—for instance, CRM agents extract data, content agents author text, and scheduling agents manage calendar events.
Agent-to-Agent Protocol (A2A): Built to work alongside MCP, A2A lets agents dynamically discover, delegate work, and collaborate in a secure fashion.
Computer Vision and UI Automation: Agents can now interact with desktop or web interfaces—clicking buttons, filling forms, and completing workflows automatically, enhancing automation capabilities.
Model Context Protocol Integration and Shared Memory: Agents maintain user context across long workflows using the MCP and Cosmos DB–powered memory in Copilot Studio.
Security & Compliance Features: Governance is baked in via Microsoft Purview, Defender for Cloud integration, Agent ID tracking, audit logging, prompt shields, and compliance alignment with the EU AI Act.

Copilot Studio increasingly functions as a no/low-code agent factory—extending AI abilities across Microsoft 365, Dynamics, and bespoke business scenarios.

7. Microsoft 365, Teams & Copilot Tuning

7.1 Copilot Tuning in Microsoft 365

Microsoft 365 Copilot introduced Copilot Tuning, allowing organizations to build custom-tailored agents using proprietary corporate data—enriching context and tailoring output for domains like legal, sales, or finance.

The Microsoft 365 Agent SDK is now available in C#, JavaScript, and Python, enabling developers to publish private (“Agent Store”) or public Copilot agents for use within Teams, Outlook, Word, and other apps Medium.

7.2 Teams: Collaborative Agents

Microsoft Teams now supports next-level automation:

Agents that maintain conversational state and recall past interactions within Teams channels.
Multi-agent orchestration for scheduling, content generation, reminders, and workflows.
Low-code interfaces to build and deploy Teams-integrated business assistants.

8. NLWeb in Action: Democratizing Conversational Web

NLWeb empowers developers to embed natural-language AI search capabilities directly within websites and apps. Early implementations at Serious Eats, TripAdvisor, Shopify, and other sites allow users to interact via conversational queries and retrieve contextually relevant content—using only a thin protocol layer and Azure infrastructure under the hood. Microsoft aims to decentralize AI chat interfaces, making conversational access accessible and affordable for all publishers.

9. Microsoft Discovery: AI for Scientific Breakthroughs

Build 2025 also featured Microsoft Discovery, a platform designed to accelerate research using multi-agent AI. Though details remain limited, the platform appears focused on generating hypotheses, orchestrating literature reviews, and coordinating multidisciplinary pipelines—leveraging the same Foundry and Studio frameworks.

10. Security, Compliance, and Responsible Agentic AI

As AI agents gain autonomous capabilities, Microsoft emphasized a layered approach to governance, risk management, and oversight:

Prompt Shields and Data Privacy: Mechanisms to isolate user prompts, manage data labeling, and prevent malicious prompt injection.
Defender for Cloud & Purview Integration: Azure-based threat detection, data loss prevention, audit logs, and compliance workflows guard production deployments.
Agent Identity & Governance: Unique Agent IDs allow discovery, audit, and accountability. Automated documentation aligns with EU AI Act requirements.
MCP Security Research: Academic analysis highlights prompt injection vectors. Microsoft’s platform integrates runtime and design safeguards to mitigate these risks.

11. Developer Ecosystem Advances

Microsoft showcased enhancements across its development tools:

VS Code and Visual Studio: Both editors now include “Copilot Agent Mode” alongside new extensions (e.g. PostgreSQL), open‑source Copilot Chat for VS Code, and updated debugging features.
Open-Source WSL: The Windows Subsystem for Linux is now open‑source, allowing community contributions, customization, and API-level integration.
WinGet, Terminal, PowerToys: These tools have been refreshed to support AI workflows—making setup and operations easier for developers.
SQL Server 2025 (Preview): Introduces native vector search capabilities, positioning the database for AI‑centric workloads.
.NET 9 Preview: Optimized for cloud‑native and AI workloads—with improved tools and deployment patterns.
Fabric and Dataverse Updates: Integration of Cosmos DB, OneLake transformations, and a digital twins builder enhance data-first AI pipelines.
Azure Quantum: Revealed progress in topological quantum computing and quantum toolkits aimed at future AI‑backed research.

12. Industry Momentum & Competitive Context

12.1 Google vs Microsoft: Enterprise vs Consumer Focus

While Google I/O 2025 emphasized consumer AI experiences, Microsoft Build 2025 was decidedly enterprise- and infrastructure-driven. Microsoft focused on open standards, compliance, and interoperability—aiming to lead AI in regulated environments and across workplace platforms .

12.2 Collaboration with AI Ecosystem Players

Microsoft welcomed participation from OpenAI (Codex, Copilot), xAI’s Grok, Nvidia’s GPUs, and Anthropic’s MCP as part of a broader AI alliance. This collaborative stance points to a shared direction for interoperable frameworks across providers.

12.3 The Paradigm Shift: The Agentic Web

Microsoft’s narrative focused on a platform-level shift. Nadella stated we’ve moved from monolithic apps to a distributed ecosystem of intelligent agents that coordinate and execute across boundaries. With MCP and NLWeb as foundational layers, and Foundry + Copilot Studio as implementation engines, Microsoft aims to build the backbone of what it terms the agentic web.

13. Summary and Outlook

50+ AI Tools Launched: Covering developer tools, enterprise platforms, cloud and device-level deployments.
Standards to Enable the Vision: MCP and NLWeb formalize interoperable structures for agentic AI.
Core Platforms: Azure AI Foundry (cloud), Windows AI Foundry (edge), Copilot Studio (orchestration).
GitHub Copilot Evolution: From auto-complete assistant to a fully autonomous coding agent.
Enterprise Productivity: Copilot Tuning, Microsoft 365 Agent SDK, and Teams automation introduce tailored AI experiences.
Security & Governance: Built-in controls at every level ensure agent actions respect compliance needs.
Developer Tools: VS Code, Visual Studio, WSL, PowerToys and more are optimized for AI workflows.
Scientific and Domain Innovation: Microsoft Discovery points to future agentic research platforms.

Microsoft Build 2025 marks a decisive turn: AI is no longer a set of tools—it’s becoming a rich, interoperable architecture of intelligent agents that can act, coordinate, and execute on behalf of users. Whether in code, research, customer service, or enterprise workflows, Microsoft’s Agentic Web stack—rooted in open protocols and layered across cloud and edge—is being positioned as the foundation for the next era of computing.

Let me know if you’d like to dive deeper into the technical specifications of MCP, NLWeb, the security architecture of Foundry, or see detailed walkthroughs of Copilot Studio agent pipelines.

2025 年 5 月，在西雅图举行的年度微软 Build 大会上，微软发布了超过 50 款全新 AI 工具，标志着其“开放智能体网络”战略正式启航。在 CEO Satya Nadella 和 CTO Kevin Scott 的领导下，微软展示了横跨 Windows、Azure、GitHub、Microsoft 365 等平台的 AI 智能体生态系统，目标是构建能够自主行动、协作，并无缝嵌入真实工作流的 AI 智能体。

1. 愿景：从被动助手到主动智能体

本届 Build 大会传递的核心信息是：我们正迈入一个由“智能体 AI”主导的时代——这类 AI 不仅仅是回应请求，而是能主动计划、决策并执行任务。Kevin Scott 指出，过去一年中 AI 智能体的使用量翻了一番，表明越来越多系统正在接管人类委托的任务。

微软提出的“开放智能体网络”类似于图形界面或互联网等计算平台转型。这一理念的基础是互操作性——微软引入了两个开放标准：MCP（Model Context Protocol）和 NLWeb，旨在让 AI 能在不同平台间无缝协作。

2. 两大开放协议：MCP 与 NLWeb

2.1 模型上下文协议（MCP）

MCP 是一个基于 JSON-RPC 的协议框架，允许 AI 智能体带着上下文与工具、数据源、安全权限协同工作。微软宣布已在其所有智能平台中集成 MCP，包括 Azure AI Foundry、Copilot Studio、Dynamics 365、GitHub、Semantic Kernel 以及 Windows 11。

MCP 能让 AI 访问本地文件系统、云数据、系统服务等，同时保障安全性与用户隐私。

2.2 自然语言网络（NLWeb）

NLWeb 由 RSS 和 Schema.org 的创建者 Ramanathan Guha 设计，允许网站开发者快速集成自然语言交互接口。它不依赖大型中心化模型，而是基于站点自身内容和向量数据库构建个性化对话体验。微软希望借此为所有网站提供“类 ChatGPT”体验。

3. GitHub Copilot：从代码助手到工程智能体

GitHub Copilot 不再只是代码提示工具，而是演变为具备“端到端开发能力”的工程智能体。它可以：

自动处理问题、生成代码、修复 Bug、编写单元测试；
与 Azure 的 SRE（站点可靠性工程）智能体结合，实现系统状态监控与自愈；
与 Copilot Studio 联动，创建数据拉取、内容生成、日程安排等多智能体工作流。

Copilot 正从助手角色转变为“可信任的工程搭档”。

4. Azure AI Foundry：企业智能体的引擎

Azure AI Foundry 是微软面向企业的旗舰平台，用于构建、部署与管理大规模 AI 智能体系统。

4.1 主要功能

支持超 1900 个模型，包括微软自研模型及第三方模型（如 xAI 的 Grok 3）；
Agent Service 正式发布，支持企业级多智能体工作流与 MCP、A2A 协议；
引入模型路由器、性能排行榜、代码解释器等组件，提升效率；
提供观测性工具（Observability）用于监控安全、成本与运行状态；
Azure AI Search 支持“智能检索”，整合对话历史与规划能力。

4.2 本地运行能力（Foundry Local）

Foundry Local 支持 Windows 与 macOS，在终端设备上直接运行 AI 模型，提升隐私性与响应速度，降低对云的依赖。

5. Windows AI Foundry 与 MCP 本地支持

Windows AI Foundry 是 Azure Foundry 的桌面扩展，提供本地模型推理能力与 MCP 注册机制：

Windows 可作为完整 AI 平台运行本地模型；
支持 MCP 注册服务、权限控制与安全沙箱；
新增 VBS Enclave SDK、后量子加密、App Actions 等安全机制；
WSL（Windows 子系统 for Linux）正式开源；
PowerToys、WinGet、Terminal 等开发工具也做了 AI 优化。

6. Copilot Studio：低代码智能体编排平台

Copilot Studio 是微软面向企业和开发者的图形化智能体编排平台：

多智能体协作：定义提取数据、生成文档、管理会议等工作流；
支持 A2A 协议，实现跨智能体任务委派；
支持计算机视觉、UI 自动化、代码解释器；
引入共享内存功能，使用 Cosmos DB 保存上下文；
深度集成 Microsoft Purview、Defender for Cloud 等安全治理工具。

7. Microsoft 365 与 Teams：生产力智能体进化

7.1 Copilot Tuning 与 365 Agent SDK

企业可以使用 Copilot Tuning 对模型进行定制训练，使智能体掌握内部知识并适配具体业务流程。SDK 支持 C#、Python、JavaScript，可将自定义智能体集成至 Outlook、Word、Teams 等应用。

7.2 Teams 中的智能体协作

Teams 支持记忆上下文的智能体、跨频道沟通、日程协调、文档创建协作等功能，进一步提升团队效率。

8. NLWeb 的实战应用

NLWeb 已被 TripAdvisor、Shopify、Serious Eats 等平台部署。用户可使用自然语言完成搜索、筛选、对话等操作，而无需跳转至第三方 AI 网站。微软希望此举能降低 AI 应用门槛，让每个网站都拥有自然语言入口。

9. Microsoft Discovery：科研智能体平台

微软发布 Microsoft Discovery 平台，旨在用 AI 智能体加速科研创新，协助进行文献调研、假设生成、多学科研究协作。细节尚未完全公开。

10. 安全、合规与责任化 AI

微软强调智能体能力越强，安全机制越要全面：

使用 Prompt Shield 防御提示注入攻击；
集成 Microsoft Defender 与 Purview 实现数据防泄漏与审计；
引入 Agent ID，标记并监管所有智能体行为；
符合欧盟 AI 法案要求，实现合规治理流程；
已有学术研究指出 MCP 的潜在安全风险，微软正在平台层面加以应对。

11. 开发者生态系统升级

微软对开发者工具做了大量优化：

VS Code、Visual Studio 集成 Copilot Agent 模式；
WSL 完全开源；
WinGet、Terminal、PowerToys 增强 AI 工作流支持；
SQL Server 2025 引入向量检索功能；
.NET 9 更适配云原生和 AI 任务；
Fabric、Dataverse、Azure Quantum 等平台集成向量、数字孪生与量子能力。

12. 产业趋势与竞品分析

微软的 Build 与 Google I/O 形成鲜明对比：前者聚焦企业 AI、开源协议与基础设施，后者更偏向消费场景与模型发布。微软希望以开放策略引领 AI 工作流标准制定。

xAI（Grok）、OpenAI、Nvidia 等公司均参与 Build 活动，展示出 AI 行业逐渐走向开放协同。

Satya Nadella 表示：“我们正从封闭的应用转向开放的智能体网络，这是一场平台级变革。”

13. 总结与展望

微软发布超 50 款 AI 工具；
打造 Azure Foundry、Windows Foundry、Copilot Studio 三位一体的智能体堆栈；
开放协议 MCP 与 NLWeb 构建统一标准；
GitHub Copilot 变身工程智能体；
企业定制智能体与 Microsoft 365 深度融合；
安全、合规能力内建；
科研、教育、金融等行业进入 agentic AI 实践阶段。

微软正在推动 AI 从“工具”进化为“系统”，并通过开放协议、标准化平台与端云协同，构建新一代智能体网络的基础架构。

注：本文内容和插图由ChatGPT生成。