Microsoft Build 2026 Developer Preview: Windows Agent Runtime, Multi-Model Copilot, and the Agentic Stack

Windows Becomes an Agent Platform

The structurally most significant announcement at Build 2026 is a platform designation, not a product launch. Microsoft is formally repositioning Windows as the execution environment for AI agents at the operating system level — not just a desktop that runs AI applications, but a sandboxed runtime that treats agents as first-class system constructs with capability grants, lifecycle management, and a distribution channel analogous to the Microsoft Store.

Windows Agent Runtime (Preview)

The Windows Agent Runtime preview ships to Windows Insiders in June 2026 alongside the Build announcements. The initial preview scope is intentionally narrow: text-based agents operating on structured data formats, specifically JSON, XML, and PDF files. Vision-based agents that can see and interact with UI elements are on the roadmap for 2027. The sandboxing model issues per-agent capability grants covering file system scope, network access, and application launch permissions. Users review and approve capability requests during agent installation, analogous to the permission model for mobile applications.

Windows Agent Store

The Windows Agent Store is the distribution and discovery layer for Windows Agent Runtime agents. Publishers submit agents through a Microsoft security review covering capability disclosure, data handling policy declaration, and sandboxing compliance verification. The store launches in preview alongside the Windows Agent Runtime in June. For developers building agents intended for commercial distribution, the security review requirements are published in the Windows Agent Runtime SDK documentation and closely mirror the Microsoft Store certification process.

AI Foundry for Windows SDK

Microsoft is shipping a new NuGet package — AI Foundry for Windows — that bundles the ONNX Runtime, DirectML, and the Copilot Runtime into a single installation target. This resolves a fragmentation problem that has made on-device AI development on Windows unnecessarily complex: applications wanting local model inference, DirectML GPU acceleration, and Copilot API access previously required three separate SDK installations with incompatible versioning chains. The unified package ships as a stable release at Build, making on-device inference a first-class deployment option for .NET developers without environment management overhead.

DevUI: Agent Execution Debugger

DevUI is a browser-based local debugger shipping in preview at Build that visualizes agent execution in real time. It shows which agent is active in an orchestrated run, what memory state the agent is drawing on, how it is routing work to subagents, and what tool calls are pending or failed. The interface is similar to a distributed trace profiler but built specifically for agent execution graphs. Teams debugging multi-agent production failures have historically worked from Application Insights logs and execution trace dumps. DevUI makes the execution model visible without requiring custom instrumentation code.

Multi-Model Copilot and Azure AI Foundry

Microsoft’s decision to rebuild Copilot as a multi-model platform is one of the most commercially significant AI architecture choices the company has made since the original OpenAI partnership in 2023. The new Copilot orchestration layer routes requests across OpenAI models (GPT-5.5 and successors), Anthropic’s Claude family, and small language models optimized for on-device Windows inference. Model selection is configurable at the enterprise tenant level and, in select scenarios, at the individual user preference level.

For enterprise IT administrators, this eliminates the single-vendor model dependency that has been a procurement concern since the original Copilot for Microsoft 365 rollout. For developers building Copilot extensions and plugins, the multi-model routing is largely transparent at the request interface level — the same API surface handles all model variants regardless of which provider performs the inference. Azure AI Foundry acts as the routing and governance layer, with per-tenant model usage logged for compliance auditing.

Azure AI Foundry also receives the Foundry Agents Service GA at Build — the managed service layer that wraps the Microsoft Agent Framework with production SLAs, auto-scaling, and a pre-built observability stack. Teams who want Agent Framework capabilities without managing agent infrastructure directly should evaluate Foundry Agents Service pricing against direct deployment to Azure Container Apps for their expected invocation volumes. The Microsoft Agent 365 complete guide covers the architectural decision points for enterprise deployments in detail.

GitHub Copilot: Autonomous Coding at Scale

GitHub Copilot’s autonomous coding agent — which can read issues, write implementation code, create tests, and open pull requests without developer prompting — was Build 2025’s headlining demo. At Build 2026, Microsoft is expected to present what one year of production deployment across real codebases has taught the team, including where autonomous agents consistently succeed (well-scoped refactors, test generation, documentation updates) and where they require closer supervision (architecture decisions, security-sensitive paths, cross-repository dependency management).

The anticipated focus areas for the Copilot announcements are multi-agent workflows inside VS Code, where separate specialized agents handle linting, test generation, documentation, and security review in parallel rather than sequentially; deeper GitHub and Azure AI Foundry integration that connects repository events directly to orchestrated agent runs; and updates on the June 1 AI Credits billing transition that changes how Copilot usage is metered for enterprise Microsoft 365 customers. Any team managing Copilot for a group of developers should review the AI Credits transition documentation before the June 1 cutover to have an accurate cost model from day one.

What Developers Should Do Before June 2

If you are building on Microsoft’s stack and want to be positioned to evaluate the Agent Framework on day one of the GA release, there are three concrete steps worth taking before the keynote:

1. Install the Microsoft Agent Framework preview NuGet package or Python wheel. Running through the quickstart samples before the GA announcement gives you a comparison baseline for evaluating what changed between preview and production, and which patterns to carry forward versus rebuild.
2. Audit your Copilot Studio agents for A2A compatibility. Agents built before the May 2026 Copilot Studio update use the previous message format for agent coordination. If you have existing multi-step automations, mapping the agent boundaries and communication patterns now makes the migration to the GA A2A protocol a scoped, planned effort rather than a reactive one.
3. Review the June 1 AI Credits billing structure before it takes effect. The transition from per-seat Copilot licensing to per-invocation AI Credits changes the cost equation significantly for high-volume deployments. Running a usage audit against current Copilot integration telemetry before the cutover is the only way to avoid billing surprises in the first week of the new model.

For teams also evaluating where Microsoft’s agentic platform fits against other providers — particularly Anthropic’s Claude Code and Google’s Gemini Enterprise stack — the A2A interoperability introduced by the May Copilot Studio update means agents from all three providers can now share tool surfaces through MCP. That is a materially different integration landscape than existed six months ago, and it changes the calculus of building on any single provider’s orchestration layer.

The Strategic Read

Microsoft Build 2026 is the conference where Microsoft publicly commits to an architectural bet it has been building toward since mid-2025: that enterprise software’s next decade is agent orchestration platforms with application logic embedded, not SaaS applications with AI features added. The Windows Agent Runtime, the unified Agent Framework, the multi-model Copilot layer, and the Windows Agent Store are four surfaces of a single architecture — not four separate product launches.

Developers who internalize that coherence before the June 2 keynote will be better positioned to navigate the announcement volume and focus on the two or three changes relevant to their current stack. The full session catalog and documentation will publish to the Microsoft Developer YouTube channel and the official Build website simultaneously with the keynote. This guide will be updated with confirmed specifications and pricing after the June 2 keynote.