Microsoft wrapped its Build 2026 developer conference on June 3 in San Francisco with a stark message: the company is no longer just building AI tools—it’s building an entire platform for agents that spans every layer of its stack. The two-day event, also streamed online, reframed Windows, Microsoft 365, GitHub, Azure, and a new family of in-house AI models as a vertically integrated fabric designed for an “autonomous age.”

For years, Microsoft positioned Copilot as an assistant. Build 2026 recast it as an operating system for autonomous software—processes that can reason, plan, and act across applications, services, and devices without human hand-holding. The headline wasn’t a single product but a connective tissue: an agent runtime that lets third-party developers, enterprise IT, and end users deploy AI agents that work across the entire Microsoft ecosystem.

The most revealing slide in CEO Satya Nadella’s keynote wasn’t a demo. It was a diagram that placed Windows at the bottom, Azure as the cloud backbone, Microsoft 365 as the productivity layer, and GitHub Copilot as the developer experience—all bound together by a shared agent framework. “This isn’t about bolting a chatbot onto Word,” Nadella said. “We’re redefining what an operating system does. It schedules prompts today; tomorrow it schedules entire business processes.”

That vision hinges on a new runtime codenamed Project Orchard, which was shown publicly for the first time. Orchard acts as a universal agent host. It can run on a Windows 11 PC, on Azure container instances, or inside a Microsoft 365 tenant. An agent built for Excel can pull data from a SQL database on Azure, pass it to a Copilot in Teams, and then trigger a GitHub Actions workflow—all orchestrated by Orchard. Microsoft says the runtime understands the security context of user, device, and tenant, so it can enforce policies from Microsoft Purview and Entra ID without extra code from the developer.

Developers at the conference got their hands on a new SDK that abstracts away much of the plumbing. The Agent Framework SDK, now in private preview, provides a unified way to define an agent’s skills, memory, and identity. A “skill” can be a REST API, a local executable, or a natural-language instruction. The SDK then handles prompt grounding, tool calling, and state management. In one session, a Microsoft engineer built a travel-expense agent that read receipts from OneDrive, matched them against calendar entries in Outlook, and filed an expense report in Dynamics 365—all from a single C# file under 200 lines of code.

That simplicity is strategic. Microsoft wants enterprise developers to treat agents as just another workload. “If you can write a CRUD app, you can build an agent,” said Scott Guthrie, executive vice president of Cloud and AI. The SDK plugs directly into Visual Studio and VS Code via GitHub Copilot extensions, so developers can scaffold an agent with a natural-language prompt. Behind the scenes, Copilot generates the code, provisions the necessary Azure resources, and registers the agent in a new catalog that will eventually surface inside Microsoft 365 and Windows.

That catalog is called the Agent Store. It will be built directly into the Microsoft 365 app launcher and the Windows 11 taskbar. Users will be able to browse agents built by independent software vendors, system integrators, or their own IT department. Each agent carries a security attestation score—a measure of what permissions it needs and whether those align with company policy. Early partners include SAP, ServiceNow, and Adobe, all of which demonstrated agents that can hop between their own cloud backends and Microsoft’s productivity apps without custom middleware.

For Windows specifically, the agent story marks the biggest platform shift since the introduction of the Microsoft Store. The operating system now includes an “Agent Host” service that runs in the background, managing local agents that can interact with the file system, accessibility APIs, and hardware sensors. A manufacturing agent shown onstage monitored a factory-floor camera feed locally on a Windows IoT device, detected a defect, and then used a Teams agent to alert a plant manager—all without round-tripping video to the cloud. “Edge reasoning changes the economics,” Nadella said. “Not every millisecond of video needs to touch Azure.”

Microsoft 365 received an agent-powered overhaul that goes far beyond last year’s Copilot for Microsoft 365. The suite now exposes every major application—Word, Excel, PowerPoint, Outlook, Teams—as an agent target. That means a third-party agent can read and write documents, send emails, and join meetings as a participant with its own identity. Microsoft calls these “declarative agents” because they declare their intent and capabilities in a manifest file. An IT admin can provision a declarative agent to thousands of users with a single policy, and users can customize its behavior with natural-language instructions that stay within the tenant’s compliance boundary.

On the model side, Microsoft announced a new family of in-house large language models called IQ-2. These models, trained partially on Azure’s custom silicon, are designed for reasoning-heavy agent workloads. They come in three sizes—IQ-2 Mini, IQ-2 Plus, and IQ-2 Pro—and will be available via Azure AI Foundry and GitHub Models. Microsoft claims IQ-2 Plus outperforms GPT-4 Turbo on agent benchmark tasks that require multi-step reasoning and tool use, while costing significantly less to serve. The models also power a new “deep reasoning” mode inside Copilot that lets it break complex tasks into sub-goals and recover from errors automatically.

GitHub Copilot, already a revenue engine, gets a new “agent mode” that turns the coding assistant into an autonomous software engineer. In a demo that drew applause, Copilot took a GitHub issue written in plain English, generated a pull request, wrote unit tests, and modified a CI/CD pipeline—all while a human developer watched and approved each step. The feature will roll out first to enterprise customers in July, with a public preview later in the year. GitHub CEO Thomas Dohmke said the goal is to let developers “describe the problem, not the solution.”

Security was a constant subtext. Microsoft knows that agents with broad permissions are a dream for attackers. The company introduced a new capability called Agent Guardrails, which uses a combination of static analysis, runtime monitoring, and user-behavior analytics to detect when an agent is behaving anomalously. If an agent suddenly tries to access a SharePoint site it has never touched or starts exporting large volumes of email, Guardrails can automatically revoke its credentials and alert the SOC. The feature works across Windows, Azure, and Microsoft 365, and it’s part of the Microsoft Defender for Cloud suite.

The announcements at Build 2026 make clear that Microsoft doesn’t see agents as a separate product category. They’re the connective tissue of the company’s entire portfolio. For enterprises, that means a future where every line-of-business application can be augmented with autonomous reasoning. For developers, it means a new kind of platform play—one where writing software is less about assembling UIs and more about composing agent skills. And for Windows enthusiasts, it signals that the operating system is no longer just a launchpad for desktop apps but a runtime for intelligent, cross-application automation.

The question now is how fast enterprises will adopt this vision. Agent orchestration introduces new complexity in testing, monitoring, and compliance. Microsoft is betting that its end-to-end control—from silicon to cloud to client—will make it the default choice for companies that already run on its stack. But competitors are not standing still. Google’s Gemini ecosystem and OpenAI’s own agent APIs are racing to similar endpoints. What sets Microsoft apart is the depth of its integration: an agent written with the new SDK can run on a factory-floor PC and also sit in a Teams chat, all with the same identity and policy envelope.

For those who have watched Microsoft evolve from a Windows company to a cloud company to an AI company, Build 2026 was a watershed. It wasn’t about a new Windows version or a new Office feature. It was about turning the entire Microsoft graph into a canvas for autonomous software. As one engineer in the audience put it, “We came expecting API updates. We left realizing the OS itself is now an agent.”