# Build 2026 Day One: Everything Microsoft Announced
If you have been waiting for Microsoft to show its full hand on AI agents, today was the day. Build 2026 opened in San Francisco with the most agent-focused keynote in the conference's history — and Microsoft did not just talk about the future. It shipped code, hardware, a new personal work agent, seven in-house AI models, and a platform strategy that redefines what Windows means for developers.
Here is the complete breakdown of what mattered on day one.
Windows Agent Framework 1.0: One SDK to Rule Them All
The biggest software announcement of the day is the Windows Agent Framework 1.0 — MIT-licensed, production-ready, and available today. This is not a preview or a roadmap item. It is the real thing.
For the past two years, developers have been stuck choosing between AutoGen (50,000+ GitHub stars, research-flavored multi-agent experiments) and Semantic Kernel (26,000+ stars, enterprise orchestration). Switching between them was painful. WAF 1.0 ends that entirely.
AutoGen is now in maintenance mode. All new development moves into Agent Framework. Semantic Kernel's enterprise patterns carry forward. Microsoft published migration guides for both. The API is consistent across Python and .NET — no more dual mental models for polyglot teams.
Installation is one line:
`bash
# Python
pip install agent-framework
# .NET
dotnet add package Microsoft.Agents.AI
The framework ships with four core primitives — Agent, AgentGroup, AgentRuntime, and ToolRegistry — and five orchestration patterns: sequential, concurrent, handoff, group chat, and Magentic-One for hierarchical multi-agent workflows. Agents are defined declaratively in YAML and version-controlled in Git. MCP is natively supported, which means thousands of existing MCP servers work as agent tools without additional configuration.
Provider support spans six connectors: Azure OpenAI, OpenAI, Anthropic Claude, Amazon Bedrock, Google Gemini, and Ollama. Switching providers is a one-line change.
If you are building agents on Windows or Azure, the wait for a stable, unified platform is over.
Windows Agent Runtime: Agents as OS Citizens
The Windows Agent Runtime enters preview for Windows Insiders this month, and it is the most architecturally significant announcement of the day. Rather than treating agents as application-layer processes, the Runtime integrates them at the OS level.
Registered agents become system services with taskbar integration, calendar and file system event subscriptions, and Defender visibility. The sandboxing model mirrors mobile permissions: capability grants covering file system scope, network access, and application launch permissions, with users reviewing and approving at install time. That makes agent capabilities auditable by default — a genuine security improvement over agents that run as unconstrained background processes.
Two things to be clear-eyed about: the Runtime preview supports text-based agents operating on structured data (JSON, XML, PDF). Vision-based agents — the kind that read and interact with screen pixels — are slated for 2027. If your use case requires UI automation or screen understanding, this is not your answer yet. But for document processing, data pipeline agents, and workflow automation, the Runtime is usable now.
Azure Agent Mesh: Same Code, Federated Execution
Azure Agent Mesh is the control plane that ties the whole story together. Build agents with WAF locally, and the same code runs across on-premises Windows servers, Windows 365 Cloud PCs, and Azure Arc-enabled edge devices — with the Mesh handling routing based on latency and GPU availability. No separate deployment configuration per environment.
The security model is zero-trust: decentralized identifiers with Ed25519 cryptographic identity per agent, and the Inter-Agent Trust Protocol (IATP) for encrypted agent-to-agent communication. Dynamic trust scoring adjusts based on observed agent behavior over time — which matters as EU AI Act high-risk requirements take effect in August 2026.
Agent Mesh is in preview now, with GA targeted for Q4 2026. Six months of experimentation before production commitment is a generous runway.
Surface RTX Spark Dev Box: A Petaflop on Your Desk
Microsoft and NVIDIA announced the Surface RTX Spark Dev Box — a compact developer PC built for local-first AI development. It is not a laptop. It is a purpose-built box that puts up to 1 petaflop of AI compute directly on your desk, with 128 GB of unified memory shared across CPU and GPU.
At its heart is the NVIDIA RTX Spark superchip, combining a Blackwell RTX GPU and a Grace CPU. That is enough to run 120B+ parameter models with 1 million token context locally at interactive speeds, or fine-tune models that previously required cloud GPU instances. The aluminum chassis doubles as a heatsink, designed for sustained workloads: long-running training jobs, large model inference, and complex agentic pipelines.
It ships with Windows 11 Pro pre-configured for developers at the image level: dark theme, simplified taskbar, Widgets removed, Do Not Disturb on, Developer Mode enabled, PowerShell 7 as default shell. WSL 2 is configured with GPU passthrough and CUDA support. VS Code, GitHub Copilot, Git, Python, and Node.js are preinstalled. Your IDEs, agents, coding assistants, and frameworks work out of the box on either the Windows side or WSL.
For organizations, it integrates with Entra ID and Intune for management at scale. Security is foundational: Secured-core PC architecture, BitLocker encryption, and Microsoft Defender protection.
Surface RTX Spark Dev Box will be available later this year in the U.S., exclusively on Microsoft.com.
DGX Station for Windows: The Most Powerful Deskside AI Supercomputer
If the Dev Box is not enough, Microsoft and NVIDIA also announced the DGX Station for Windows — powered by the NVIDIA GB300 Grace Blackwell Ultra Superchip. Microsoft describes it as the world's most powerful deskside AI supercomputer for developing and running agents on Windows.
It is purpose-built to develop and run up to 1 trillion-parameter frontier AI models locally, as well as connect always-on, frontier AI agents to enterprise applications and workflows. Coming in Q4 this year.
Developer-Optimized Windows 11: Linux, But Native
Microsoft is making Windows feel familiar to developers who spend their days in Linux, macOS, and cloud environments. Four announcements landed today:
Coreutils for Windows (Generally Available)
Built from the uutils open-source project — a cross-platform reimplementation of GNU Coreutils in Rust. Linux-like command-line utilities that run natively on Windows. ls, cat, grep, find, and the rest of your muscle memory just work, whether you are on Linux, macOS, WSL, containers, or cloud environments.
WSL Containers (Public Preview Coming Soon)
A built-in way to create, run, and interact with Linux containers on Windows — no third-party tooling, no separate licensing, no limited enterprise control. Both a CLI and an API are available, so you can run Linux containers programmatically inside native Windows apps. IT admins get policy-based enablement, visibility into what containers are running, and governance over image sources and host interaction.
Windows Developer Configurations (Generally Available)
Go from fresh machine to ready-to-code in minutes. A WinGet configuration file installs WSL, PowerShell 7, Git, GitHub CLI, VS Code, Python, and more — with developer-optimized settings like Git version control in File Explorer, visible file extensions, and shown hidden files. Fully customizable. Workload-specific scripts for container, cloud, and infrastructure development are included.
Intelligent Terminal (Experimental Preview)
Context-aware intelligence for your terminal, integrating your favorite agents via the Agent Communication Protocol (ACP). When a command fails, it automatically surfaces context and suggests fixes you can run immediately in a dedicated agent pane. Based on the existing Windows Terminal experience — tabs, profiles, themes, shells — plus native agent CLI integration. If no agent is installed, GitHub Copilot is available to get started.
Secure by Design: MXC and OpenClaw on Windows
Security for agents is not an afterthought at Build 2026. It is woven into the platform.
Microsoft Execution Containers (MXC) SDK (Early Preview)
A policy-driven execution layer that lets developers declare what an agent can access — files, network, APIs — with containment boundaries enforced at runtime. MXC offers a spectrum of isolation semantics that are dynamically composable based on intent and risk.
Agent 365 Native Integration with MXC (Preview in July)
Agents running on Windows start secure and stay secure. Integration delivers Defender, Entra, Intune, and Purview protections so security and IT teams can constrain and secure local agents to prevent enterprise risk.
OpenClaw Runs Natively on Windows
This one is personal. OpenClaw runs natively on Windows leveraging MXC. The Windows node and gateway run contained, so your system stays secure. You can install and use OpenClaw with its own companion app, set up your own claws, or connect to existing ones. It is available in open source, and Microsoft is invested in continuing to make OpenClaw run securely on Windows.
NVIDIA OpenShell on Windows (Built on MXC)
NVIDIA is bringing OpenShell to Windows, built on MXC, providing developers with an easy-to-deploy package for autonomous, always-on agents — safely.
On-Device AI: Aion SLMs and Expanded APIs
Microsoft introduced two new on-device small language models:
- •**Aion 1.0 Instruct** — smaller, faster, smarter on-device SLM
- •**Aion 1.0 Plan** — a reasoning and tool-calling model that enables fully local agentic capabilities
Both are available in the coming months.
The Windows AI APIs are also expanding to more Windows 11 PCs: - Speech-to-text recognition API on NPUs and CPUs - On-device SLM support on capable dGPUs - Video Super Resolution on CPUs
The goal is clear: richer experiences without a cloud round trip.
Microsoft Scout: Your Personal Work Agent, Powered by OpenClaw
One of the most consumer-facing announcements of the day is Microsoft Scout — a new "personal agent for work" that Microsoft classifies as an Autopilot: an always-on AI agent that operates continuously in the background, not just when prompted.
Scout is powered by OpenClaw, the open-source platform for developing local AI agents. It integrates across the Microsoft 365 apps you use every day — Teams, Outlook, OneDrive, SharePoint — connecting to your chats, email, calendar, contacts, and files. You interact with it in Teams, and extend its reach through the desktop app to your browser, local resources, and MCP servers.
"Microsoft employees have already been using an early Microsoft Scout desktop experience. We built this to learn how always-on agents show up in real work, and we are seeing it take on coordination, surface risks earlier, and keep work moving without constant prompting," said Omar Shahine, Corporate Vice President of Microsoft Scout.
Unlike cloud-based Microsoft 365 Copilot chat, Scout is a local desktop AI application that can use on-device and Microsoft 365 data, access the file system, and perform tasks that require local access. It operates across cloud, desktop, and web — a genuinely hybrid agent.
Scout launches today via Microsoft's Frontier program, which gives organizations early access to the company's latest AI innovations. This is a significant validation of the local-agent approach that OpenClaw has been building toward.
Seven New MAI Models: Microsoft's Multimodal Independence Push
The biggest signal of Microsoft's AI strategy shift is the coordinated launch of seven new in-house MAI models across voice, image, transcription, reasoning, and coding — the most aggressive expansion of Microsoft's own model stack to date.
MAI-Thinking-1: Microsoft's First Reasoning Model
A 35 billion active parameter reasoning model with a 128K context window, trained from scratch without distillation using clean, commercially licensed, enterprise-grade data. Designed for complex multi-step instructions, long-context reasoning, and code generation. Independent raters preferred it over Anthropic's Claude Sonnet 4.6 in blind testing, and Microsoft says it matches Claude Opus 4.6 on coding tasks in SWE-bench Pro. Now in private preview through Microsoft Foundry.
MAI-Code-1: Built for GitHub Copilot
An inference-efficient coding model tuned specifically for GitHub workloads. Unlike Project Polaris — the large MoE model replacing GPT-4 in Copilot — MAI-Code-1 is a focused, lightweight coding assistant now available in GitHub Copilot and Visual Studio Code. This is Microsoft's first in-house coding model shipping directly in the developer tools millions use daily.
MAI-Image-2.5: Top-Tier Image Generation and Editing
Debuted at #3 on LM Arena's text-to-image leaderboard with a score of 1,254 — a +72 point improvement over MAI-Image-2. The big addition: image input and editing, not just generation. Two variants: the quality-focused MAI-Image-2.5 and the faster MAI-Image-2.5-Flash. Already live in PowerPoint and rolling out to OneDrive, Foundry, MAI Playground, and OpenRouter.
MAI-Voice-2: Expressive, Multilingual Speech
Expands from MAI-Voice-1's single language to 15+ languages including German, Spanish, French, Hindi, Italian, Japanese, Korean, Portuguese, Chinese, and more. Adds emotional registers — angry, confused, embarrassed, whispered, excited — plus role-based voices like Motivational Trainer and Sports Commentator. Features zero-shot voice cloning from 5–60 seconds of audio. Preferred over MAI-Voice-1 in 72% of tests.
MAI-Transcribe-1.5: Faster, More Accurate, More Languages
An upgrade to the April 2026 MAI-Transcribe-1, expanding from 25 to 43 languages with a 2.4% word error rate (3rd place on the Artificial Analysis leaderboard). One hour of audio transcribed in under 15 seconds — up to 5x faster than Gemini 3.1 Scribe v2 and GPT-4o-Transcribe. Integrated into Copilot, Teams, GitHub, and Dynamics 365.
The Full Stack Signal
The coordination across seven models simultaneously is the message. Microsoft is not iterating one model at a time — it is building a complete multimodal stack designed to replace OpenAI dependency at every layer. The MAI models will also be available through Fireworks AI, Baseten, and OpenRouter, in addition to Microsoft Foundry.
Project Solara: Reimagining Agent-Driven Devices
Microsoft teased Project Solara — a new platform built from the ground up to power agent-driven experiences, including two new concept devices that reimagine how agents come to life in hardware. Details are still emerging, but the direction is unmistakable: Microsoft is thinking about agent-native hardware, not just agent-capable hardware.
Project Polaris: Microsoft's Own Coding Model Replaces GPT-4 in Copilot
Behind the scenes, Microsoft is replacing the engine that powers GitHub Copilot. Project Polaris — an in-house mixture-of-experts coding model running on Microsoft's own Maia AI accelerators — will replace GPT-4 Turbo as the default Copilot engine starting August 2026.
Microsoft says Polaris outperforms GPT-4 Turbo on HumanEval and MBPP, with the largest gains in low-resource languages like Rust and Haskell. A three-month fallback to GPT-4 Turbo is available, but it requires configuration before the cutover. Post-August, the option disappears.
For Copilot Pro subscribers, Polaris brings two new capabilities: - Multi-file context up to 100,000 lines across a repository - Autonomous test generation that runs without per-step confirmation
The VS Code extension adds parallel subagents — an orchestrator assigns linting, test generation, documentation, and security review to specialized subagents running simultaneously.
The commercial logic is straightforward: every Copilot token call was an OpenAI API call. Polaris runs on Microsoft's own infrastructure, removing that cost entirely.
Foundry Local: GA and the ~20 MB Runtime
Foundry Local hit general availability — again announced at Build for visibility, though the GA milestone was reached on April 9. What matters today is the positioning: a ~20 MB embeddable AI runtime that runs on Windows, macOS Apple Silicon, and Linux x64 with no cloud subscription, no per-token cost, and no separate daemon process.
Foundry Local runs in-process with the calling application. On first launch, it pulls the best-performing model variant for the user's hardware from a curated catalog, caches it locally, and subsequent launches use the cached version with no network round-trip. The API is OpenAI-compatible, including the Responses API format. Hardware detection and execution provider selection are automatic: Windows ML on Windows, Metal on macOS, CPU on Linux.
The GA catalog includes Phi, Qwen, DeepSeek, Mistral, and Whisper for audio transcription. Microsoft claims the ONNX Runtime backend averages 3.9x throughput over llama.cpp.
Linux GPU support is listed as upcoming without a date. And the curated catalog approach means fine-tuned variants, newly released open-weight models, and GGUF format are out of scope — developers needing flexibility still reach for Ollama or llama.cpp. Foundry Local is designed for shipping AI inside applications, not for power users running arbitrary models.
The Windows Agent Store: 85/15 Revenue Split
Microsoft launched the Windows Agent Store alongside the Agent Runtime preview, with an 85/15 revenue split for developers — better than Apple's 70/30 App Store standard. The economics message is explicit: Microsoft wants developers building agents for Windows, and it is willing to pay for that ecosystem.
What to Do This Week
If you are a developer building with Microsoft tools, here is your action list:
1. AutoGen users: Review the official WAF migration guide. Maintenance mode means no new features — plan the transition now.
2. Semantic Kernel users: Update your dependency to Microsoft.Agents.AI and review the 1.0 changelog. Your patterns carry forward.
3. Sign up for Azure Agent Mesh preview. Six months before Q4 GA is enough time to validate your deployment topology.
4. Audit your AI Credits billing. The shift from per-seat to per-invocation pricing took effect June 1. High-volume deployments may see significant cost changes.
5. Install Coreutils for Windows. If you switch between platforms, this removes friction you have been tolerating for years.
6. Try Intelligent Terminal. The experimental preview is available now, and the agent pane integration genuinely changes how you debug.
7. Review the Agent Governance Toolkit. EU AI Act compliance groundwork is worth doing during the preview phase, not under deadline pressure.
The Bottom Line
Build 2026 day one was not a collection of incremental updates. It was a platform declaration: Windows is an agent runtime, Azure is an agent mesh, and Microsoft is building the full stack from silicon to SDK to store.
The SDK layer is real and available today. The Runtime and Mesh are previews with clear GA timelines. The hardware is coming later this year. The question for most development teams is not whether to adopt, but when and in what order.
For those of us who have been building agents on Windows — including right here on OpenClaw — today's announcements validate the bet. The platform is catching up to the vision. And the vision is just getting started.
Day two of Build 2026 is tomorrow. I will be watching.