Microsoft transformed its annual Build developer conference in San Francisco this week into a launchpad for the agentic era, unveiling an ambitious platform designed to make AI agents the default building block of the next-generation web. The company extended the vision it first sketched at Build 2025, pairing new cloud, web, Microsoft 365, Windows, GitHub, and Surface developer tools into a cohesive fabric that it calls the “agentic web.”

In a two-hour keynote, CEO Satya Nadella and a parade of technical leaders demonstrated how autonomous agents, backed by a standardized communication protocol and deeply woven into every major Microsoft product, will shift software from static tools to proactive, collaborative partners. “We’re moving from a world where you work with software to a world where software works with you,” Nadella said.

The Agentic Web: Beyond Chatbots

The term “agentic web” signals Microsoft’s belief that the web itself must evolve to support persistent, goal-oriented AI entities that operate across sites, apps, and services on behalf of users. Unlike today’s chatbots that wait for prompts, these agents can plan multi-step tasks, negotiate with other agents, and execute actions while respecting user permissions. The company showcased agents that can book travel across multiple airline sites, negotiate service contracts by interfacing with supplier APIs, and manage supply chains by autonomously triggering purchase orders when inventory dips.

At the core of this vision lies the Model Context Protocol (MCP), an open standard that Microsoft championed at Build 2025 and has now submitted to the Internet Engineering Task Force for broader adoption. MCP defines how AI agents exchange context, credentials, and goals in a secure, decoupled manner. This year, Microsoft announced that MCP will ship natively in Azure, Windows, Edge, and Microsoft 365, making it the default handshake for both first- and third-party agents. “We believe MCP will do for agent interoperability what HTTP did for documents,” said Kevin Scott, CTO.

Cloud and Developer Tools: A Factory for Agents

Microsoft unveiled a major update to Azure AI Foundry, its end-to-end platform for building, testing, and deploying AI agents. The new Agent Runtime provides a serverless execution environment with built-in security sandboxes, memory management, and tool-calling capabilities. Developers can now define agents using a declarative YAML schema that’s portable across clouds and local environments — a push to avoid vendor lock-in.

Key new services include:

  • Agent Mesh: A peer-to-peer overlay network that lets agents discover each other and collaborate without a central orchestrator. It’s designed for millisecond-scale communication and runs on the same infrastructure as Azure’s global CDN.
  • Trust Boundary Manager: A policy engine that enforces user-defined data boundaries. For example, a healthcare agent can only access patient records from HIPAA-compliant regions while still coordinating with a calendar agent running on consumer infrastructure.
  • Agent Store: A marketplace where enterprises and independent developers can publish, monetize, and consume pre-built agents. Microsoft will take a 15% commission, mirroring its model for Visual Studio Code extensions.

Microsoft also integrated MCP deeply into GitHub Copilot. Starting in June 2026, Copilot will act as an agentic coding partner that understands entire repositories, can propose multi-file refactors, and even autonomously fix sprint bugs assigned in GitHub Issues. The new @agent command in the Copilot Chat interface allows developers to invoke specialized sub-agents — one for security audits, one for accessibility compliance, one for performance profiling — that collaborate using MCP.

Windows AI: The OS Becomes a Host

The next major update to Windows 11, due in the second half of 2026, will include Windows Agent Runtime, a system-level service that can run lightweight agents locally using the NPU (Neural Processing Unit) on Copilot+ PCs. This means personal agents — such as one that organizes your desktop, or one that watches for phishing attempts in real-time — can operate with minimal latency and full data privacy, no cloud round-trip required.

Windows 11 will also get Agent Hub, a centralized control panel where users can view all active agents, see what permissions they hold, and audit their actions. The hub includes a “session replay” feature that shows a step-by-step visual log of every decision an agent made, akin to browser history but for agent activity. Microsoft demonstrated an agent that automatically redacts sensitive text in screenshots before they reach the clipboard — and guarantees that processing happens entirely on-device.

Pavan Davuluri, head of Windows + Devices, emphasized that trust is the platform’s foundation. “Every agent on Windows will operate with clear, user-controlled boundaries. No agent can access the microphone, camera, or file system without an explicit, just-in-time consent prompt.”

Microsoft 365: Agents in the Flow of Work

In Microsoft 365, the agentic transformation is manifest in Business Process Agents, a new category of AI workers that can manage routine, multi-party workflows. Using Copilot Studio, line-of-business experts can now model these agents with a visual, low-code designer and embed them directly within Teams channels, Outlook groups, and SharePoint sites.

During the keynote, a live demo showed a “Supplier Onboarding Agent” that automatically collected W-9 forms from a new vendor, verified the tax ID against IRS databases, routed the document for internal approval, and finally sent a welcome email — all without a human in the loop. If an exception occurred, the agent escalated to a designated manager and provided a clear summary of the blocker.

Microsoft Graph connectors ensure that these agents have secure, scoped access to organizational data. And for the first time, agents built in Copilot Studio can now be published to the Microsoft Teams Agent Store, where employees can discover and install them much like they do apps today.

GitHub and the Agentic DevOps Pipeline

GitHub, already the world’s largest developer platform, goes deeper into agentic territory. GitHub Agent Stack provides a full toolkit for building, testing, and monitoring agentic software. It includes:

  • Agent Playground: A browser-based simulator where developers can test multi-agent scenarios, inject synthetic failures, and inspect MCP message traces.
  • GitHub Copilot for PRs: When a pull request is opened, a dedicated agent reviews the diff for security, performance, and style issues, then leaves inline comments just like a human reviewer. If the PR author enables auto-fix, the agent can commit suggestion corrections directly.
  • Agent CI/CD: A new runner type in GitHub Actions that understands naturally described intent. Instead of writing a YAML workflow, developers can say “run tests, deploy to staging, monitor for 10 minutes, then promote to production if all checks pass,” and the agent generates, validates, and executes the pipeline — with rollback capabilities.

Surface and the Hardware Angle

No Build would be complete without new Surface hardware, and this year Microsoft introduced the Surface Pro 11 for Business, purpose-built for agentic workloads. It features a second-gen Qualcomm Snapdragon X Elite processor with a dedicated NPU block capable of 75 TOPS (trillions of operations per second), enough to run multiple local agents concurrently. The device also packs 64GB of LPDDR6 memory and a hardware-backed Trusted Platform Module (TPM) 3.0 that enables verified credentials for agents, ensuring that a banking agent, for instance, is exactly who it claims to be.

Alongside the Pro, Microsoft previewed Surface Agent Display, a 27-inch desktop monitor with an embedded NPU and a privacy shutter that physically blocks the camera when no agent requires it. The display can sense when a user walks away and automatically lock the screen, handing active sessions to a low-power agent that continues listening for important notifications.

Developer Reactions and the Road Ahead

While the deep integration across Microsoft’s ecosystem drew applause from many attendees, some third-party developers expressed concerns about fragmentation. If Microsoft’s implementation of MCP diverges from the open standard, they fear a repeat of the embrace-extend-extinguish cycle. In response, Scott promised to release reference implementations under the MIT license and to maintain full compatibility with non-Microsoft agents.

Others worried about the sheer complexity of debugging agentic applications. Microsoft addressed this with Agent Trace, a distributed tracing tool integrated into Visual Studio, VS Code, and the Azure portal. It visualizes the entire sequence of agent-to-agent calls, prompting, and tool execution — akin to a network trace for agents.

Looking forward, Microsoft confirmed that the agentic platform will evolve through monthly updates, with a clear public roadmap to be posted on the MCP GitHub repository. The company also teased an “agentic browser” — an Edge feature that can orchestrate tasks across tabs and websites while keeping the user in the driver’s seat via a new permission model. No release date was given, but insiders suggest a preview at Build 2027.

Conclusion: A Platform for the Next Web

With Build 2026, Microsoft has painted a compelling picture of an internet where AI agents are first-class citizens. By embedding agentic capabilities into the cloud, the OS, the browser, the office suite, and even the hardware, the company is betting that developers will flock to build the next wave of intelligent applications. The unified MCP protocol, if adopted broadly, could reduce the friction that has kept agents siloed within walled gardens. Yet questions remain about governance, security, and the economic displacement such autonomous agents might bring. Microsoft’s answer, for now, is to put extensive transparency and control mechanisms front and center — and to invite the entire developer community to co-create the agentic web.