Summary
Microsoft announced Windows 11 support for Microsoft Execution Containers, which let organisations set runtime limits on what AI agents can access. The company also outlined local-model features for Copilot and new Windows PCs designed for AI workloads.
Microsoft is adding controls for AI agents and support for running more AI workloads locally in Windows, according to an October 7 announcement on its Windows Experience Blog. The changes include the general availability of Microsoft Execution Containers (MXC) on Windows 11, new local-model options for developers, and Copilot features that are expected to begin rolling out to Copilot+ PCs in the coming months.
The plan combines computing on a PC with access to cloud AI. Microsoft says Windows will route work to a local model when appropriate and use cloud services when needed. That could let some tasks run on the device while reducing reliance on cloud processing; the announcement does not specify how routing will work for each Copilot task.
Windows adds controls for agents and local models
AI agents can run tools, access files and carry out tasks with less step-by-step user involvement than a conventional app. Microsoft says MXC is designed to limit that activity: organisations can define which files and networks an agent may access, with those policies enforced while it runs. The company says MXC is now generally available on Windows 11.
Microsoft also describes a system for identifying which agent took an action and connecting agent activity to management tools such as Agent 365 and Intune. It says governance at scale may require additional services. The company lists Codex, GitHub Copilot, OpenClaw, Replit and LM Studio among agents that already support MXC, with more integrations planned.
For developers, Microsoft announced support for llama.cpp in Windows ML, its runtime for deploying models across GPUs, neural processing units and CPUs. It also says GitHub’s HydraFusion task-routing system will be able to use models running locally on Windows. That integration is due in experimental preview later in October for the GitHub Copilot app, GitHub Copilot CLI and Visual Studio Code.
Microsoft highlighted MAI Code 1.1 Flash as an example of a model intended for local coding work. It has 137 billion total parameters and 6.8 billion active parameters. The company says using 3-bit precision reduces its size by nearly 80% while retaining coding quality, and that it supports a 256K context window locally. These are company-reported model capabilities, rather than a guarantee that every Windows PC can run the model at the same speed or quality.
Copilot features and AI PCs will arrive in stages
On Copilot+ PCs, Microsoft plans to let Copilot use relevant local files and recent activity with the user’s permission, take actions across Windows, and call models running on the device. The company gives examples including organising files and troubleshooting device issues. It says the features are expected to begin rolling out over the coming months; timing may vary by device, market and silicon platform.
Microsoft also announced a simpler setup for running agents on mini desktop PCs, including a native Windows gateway for OpenClaw with MXC integration. For more demanding local workloads, it announced pre-orders for Windows PCs built around NVIDIA RTX Spark, including the Surface Laptop Ultra. Microsoft says that Surface model has up to 128 GB of unified memory and can run AI models exceeding 120 billion parameters locally. The RTX Spark PCs listed in the announcement are due to begin shipping on October 16. The Surface RTX Spark Dev Box is due to ship to customers in the United States in November.
At the higher end, Microsoft said Windows systems using NVIDIA DGX Station are planned for later this year. The company says these deskside systems will use the GB300 Grace Blackwell Ultra Desktop Superchip and support models of up to 1 trillion parameters. The announcement also included a new Windows Search experience, with taskbar actions such as changing settings or arranging windows beginning to roll out to Windows Insiders in the experimental channel.