Microsoft announced on October 7, 2026, a new direction to position Windows as a platform for “hybrid intelligence,” in which artificial intelligence tasks are executed locally on the computer when resources are sufficient, while more demanding tasks are forwarded to the cloud. The company says the growing size of models is increasing demand for usage to levels that may exceed cloud-service budgets, making the combination of the two approaches a way to reduce token consumption while retaining advanced capabilities.
A More Isolated Environment for AI Agents
Microsoft began the public availability of the Microsoft Execution Containers (MXC) environment designed to run artificial intelligence agents. The environment enables organizations to specify the files and networks that each agent can access, then enforce these policies during execution.
MXC supports several isolation methods, including process and session isolation, WSLc, virtual machines, and the Windows 365 for Agents service. Microsoft is building its governance model around three pillars: containment, identity, and manageability, with the ability to distinguish between agents and determine the actions each one performed, through integration with Microsoft Agent 365 and Intune.
Products currently compatible with it include OpenAI’s Codex, GitHub Copilot, OpenClaw, Replit, LM Studio, NVIDIA’s OpenShell, and Unsloth AI. The company also announced plans to add Anthropic’s Claude Code, Box, Manus, Perplexity, and Raycast, along with Meta’s Muse as a native Windows application.
Local Models and Automatic Task Distribution
Microsoft is working to run the MAI Code 1.1 Flash coding model locally. The model contains 137 billion parameters in total, including 6.8 billion active parameters, with its size reduced by approximately 80% through 3-bit quantization, while retaining a context window of 256,000 tokens.
Devices equipped with an RTX Spark chip will also be able to run large models locally, including NVIDIA’s Nemotron with more than 70 billion parameters after being quantized to 2 bits, and DeepSeek V4 Flash with 284 billion parameters. Microsoft will use GitHub’s HydraFusion technology to distribute tasks between cloud and local models. This capability is scheduled to be released experimentally during October in GitHub Copilot, GitHub Copilot CLI, and Visual Studio Code, with support for the open-source llama.cpp inference engine within Windows ML.
Copilot Handles Device Files and Actions
Copilot will receive local capabilities on Copilot+ PC devices, subject to the user’s approval. These capabilities include understanding files and recent activity on the device, performing actions such as organizing files, diagnosing the device, troubleshooting, and coding, as well as selecting a suitable local model or combining it with a cloud model.
The capabilities will extend to the Copilot components called Home, Code, and Autopilot. Code enables the creation of native Windows applications from a single prompt, with secure code execution through MXC, while Autopilot can continue working in the background while the user is occupied with other tasks. Microsoft said the rollout will take place gradually over the coming months.
What Changes in Practice?
The announcement represents a shift from presenting Copilot as an assistant that relies primarily on the cloud toward building an operating layer for agents capable of reading local data and performing actions on the device. The potential benefit is reducing data transmission and reliance on cloud connectivity for some tasks, but this depends on user approval, organizational policies, and hardware capabilities. Running large models locally is also not intended for all Windows devices; the examples provided are associated with computers offering substantial memory and computing resources.
In parallel, Microsoft opened reservations for the Surface Laptop Ultra computer equipped with an RTX Spark chip and up to 128 gigabytes of unified memory, saying that it can run models exceeding 120 billion parameters locally. Shipping in the United States begins on October 16, while shipping for the Surface RTX Spark Dev Box begins in November. The price of the 24-gigabyte Surface Laptop Ultra version is 513,480 yen, while the 128-gigabyte version costs 1,094,280 yen.