Microsoft positions Windows as a hybrid intelligence platform: MXC containers go GA, Copilot gains local files, local actions and local models

The first block is security, the part Microsoft details most. The post calls Windows "the most secure platform for agents," built on three primitives: containment controls what an agent can access; identity makes clear which agent acted, so IT can distinguish agent activity from the person using the device; and manageability connects that activity into tools like Agent 365 and Intune so organizations can set policy, monitor risk and govern agents. In product terms, Microsoft Execution Containers (MXC) become generally available on Windows 11 today: organizations can define which files and networks agents can access, with those policies enforced at runtime. For individuals using AI on a personal PC, Microsoft says these safeguards are part of the agent experience itself. MXC works across operating systems, but its deep Windows integration adds containment options from process and session isolation to WSLc, virtual machines and Windows 365 for Agents.

The MXC support list is worth noting on its own. Leading agents already supporting MXC: Codex from OpenAI, GitHub Copilot, OpenClaw, Replit, LM Studio, OpenShell from NVIDIA and Unsloth AI. Many others will be releasing support, including Anthropic Claude Code, Box, Egnyte, Heidi Health, Hermes Agent by Nous Research, Manus, Perplexity, Raycast and Simular. Meta's personal AI agent, Muse for Windows, is also coming as a native Windows app with MXC integration. In other words, Microsoft is not requiring anyone to use its own agent — it is making containment a platform capability first.

The second block is models and runtime. Microsoft says customers' needs are outpacing what their cloud budgets can support: they want every AI token to count without giving up frontier capabilities, so local and cloud AI need to work together. To bring frontier coding intelligence to the device, Microsoft takes MAI Code 1.1 Flash, introduced at Build as a 137 billion total and 6.8 billion active parameter model, and puts it on the PC using 3-bit precision to cut model size by nearly 80% while preserving coding quality and supporting a 256K context window locally. Alongside it, an upcoming Nemotron model from NVIDIA comes in over 70 billion parameters, quantized in 2-bit to use just over 20GB of memory, and DeepSeek V4 Flash — a 284B parameter model — plus others run locally on RTX Spark. The stated goal is to keep AI responsive while leaving room for everything else on the PC.

The third block is routing and the runtime ecosystem. Earlier this year GitHub launched HydraFusion, which routes each task to the right model in the cloud. Microsoft is now extending that intelligence to Windows, enabling HydraFusion to tap into models running locally on device so your tokens go further. Per the post, hybrid intelligence on Windows powered by HydraFusion is coming to the GitHub Copilot app, GitHub Copilot CLI and Visual Studio Code in an experimental preview later in October. On runtime, Windows ML is the scalable runtime for deploying models across GPU, NPU and CPU, and Microsoft announced llama.cpp support in Windows ML, giving developers more open-source model choices.

The fourth block is what changes for Copilot, the most directly visible piece. On Copilot+ PCs, Copilot gets supercharged with Windows hybrid intelligence and three local capabilities: local context (with your permission, Copilot can understand relevant content on your PC, including files and recent activity), local actions (Copilot can act on your behalf across Windows, helping organize files, assess device diagnostics, troubleshoot, code and carry out workflows), and local models (Copilot can use AI models running on your PC, using local compute for the right tasks). These run across the new Copilot's Home, Code and Autopilot experiences: in Home it can pull in the files you have been working on and turn them into collaboration-ready artifacts; in Code it helps create native Windows apps from a single prompt, run code more securely with MXC support and use local models where appropriate to manage token costs; and for Autopilot, a persistent personal agent, hybrid intelligence grounds the agent flow in local context, local models and PC actions. The scale figures Microsoft gives: over 2 trillion inferences run locally per month across all Copilot+ PCs, and over 40% of laptops being built for business are Copilot+ PCs.

The fifth block is devices. Microsoft announced that new RTX Spark-powered Windows PCs from ASUS, Dell, HP, Lenovo, MSI and Microsoft Surface are available for pre-order today and begin shipping October 16; RTX Spark-based dev boxes and mini desktops arrive later this year. Surface Laptop Ultra, described as the most powerful Surface ever, combines up to 128GB of unified memory with the ability to run models exceeding 120 billion parameters locally, available October 16; Surface RTX Spark Dev Box ships to customers in the US in November. On performance, Microsoft says Windows PCs with RTX Spark deliver up to 2.1x faster time to first token, 4.3x faster AI image generation and 6.2x faster AI video generation compared with a 16-inch MacBook Pro with M5 Pro. Worth stating plainly: these are Microsoft's own comparison figures, i.e. vendor-run benchmarks.