What does hybrid intelligence mean for your Windows PC?

Agents that run where the work fits

Pavan Davuluri, executive vice president of Windows and Devices, published the announcement on October 7, 2026. He describes a foundational shift. Agents are no longer chat windows. They write code, access files, and act across systems without constant supervision. That change demands a platform where compute, security, and control move together.

Hybrid intelligence is the label Microsoft uses for this platform. The idea is straightforward. Work runs locally when the hardware can handle it. Work reaches the cloud when the model or the data requires it. The routing decision happens automatically. GitHub HydraFusion, previously cloud-only, now taps models on the device. That extension arrives in experimental preview for GitHub Copilot app, GitHub Copilot CLI, and Visual Studio Code later in October.

Containers that hold the line

Traditional sandboxes were not built for agents. Containment needs shift with every prompt and tool call. Microsoft Execution Containers (MXC) address that gap. Today MXC becomes generally available on Windows 11. Organizations define which files and networks an agent may touch. Those policies are enforced at runtime.

MXC works across operating systems. Its deep Windows integration adds control options: process isolation, session isolation, WSLc, virtual machines, and Windows 365 for Agents. For individuals using AI on a personal PC, the safeguards are part of the agent experience itself. A Windows Update is required. Features roll out over time. Availability and timing vary by device, market, and silicon platform.

Leading agents already support MXC. The list includes Codex from OpenAI, GitHub Copilot, OpenClaw, Replit, LM Studio, OpenShell from NVIDIA, and Unsloth AI. More are coming. Anthropic Claude Code, Box, Egnyte, Heidi Health, Hermes Agent by Nous Research, Manus, Perplexity, Raycast, and Simular are among those preparing support. Muse for Windows, a personal AI agent from Meta, will arrive soon as a native app with MXC integration.

Models that fit on a laptop

Frontier coding intelligence is moving to local hardware. MAI Code 1.1 Flash carries 137 billion total parameters and 6.8 billion active parameters. At 3-bit precision the model shrinks by nearly 80 percent while keeping coding quality and a 256K context window on the device.

Other models join the lineup. An upcoming Nemotron model from NVIDIA exceeds 70 billion parameters. Quantized to 2-bit it occupies just over 20GB of memory. DeepSeek V4 Flash brings 284 billion parameters. The goal is to keep AI responsive while leaving room for everything else you do on your PC.

Windows ML provides the runtime. It scales across GPU, NPU, and CPU. Llama.cpp support is now announced for Windows ML. Developers gain more