Microsoft advances hybrid AI on Windows and launches new devices

Microsoft has unveiled a new phase of Windows focused on hybrid artificial intelligence, combining local and cloud processing. The package includes agent controls, integration of locally installed models with GitHub Copilot, and new Copilot features for compatible machines. These innovations have different release timelines.

Agent control with MXC

One of the new features is the general release of Microsoft Execution Containers (MXC), an execution layer that restricts resources accessed by agents and AI-generated code. Developers and administrators define which files and network destinations can be used, and the execution environment enforces these rules without the agent’s interference.

An agent cannot be its own security authority, said Logan Iyer, corporate vice president of platform and development for Windows.

For example, an agent responsible for updating a website may need to modify code and check server settings but should not have permission to change those settings.

MXC offers various isolation levels. On Windows, one option runs the agent in a separate session with its own desktop and limited interaction with the user’s session. The system also has distinct operating modes: in enforcement mode, unauthorized access is blocked; in permissive mode, it is logged but may continue.

Local models integrated with GitHub Copilot

In GitHub Copilot, programmers can allow the system to choose between local and cloud models or select a locally installed model directly. The automatic selection considers task context and prior work, aiming to balance capacity, cost, and response time.

The company notes that running a model locally does not make the entire session offline. Model selection, response processing, and tool usage have different limits. Additionally, agent commands can inherit the permissions of the executing account, while isolation adds rules about what these processes can do.

The hybrid integration with GitHub Copilot, its command-line interface, and Visual Studio Code is expected to launch in an experimental version in October.

New Copilot features on Windows

On Copilot+-compatible computers, Microsoft expects the assistant to use recent files and activities as context, with permission. Planned actions include organizing files, analyzing diagnostics, and helping resolve device issues.

The integration also extends to Autopilot, Copilot’s persistent agent. Distribution of these hybrid features is expected to begin in the coming months.

New computers support the strategy

Microsoft has started pre-selling the Surface Laptop Ultra and Surface RTX Spark Dev Box, both designed for development and local AI workloads.

The laptop offers up to 128 GB of unified memory and, according to the company, can locally run models with over 120 billion parameters.

The Surface Laptop Ultra will be available starting October 16. Shipping of the Surface RTX Spark Dev Box to U.S. customers is scheduled for November.

Frequently Asked Questions

What is hybrid intelligence? It combines models running on the computer and in the cloud, distributing work according to each task’s needs.

What does MXC control? It enforces access rules for resources like files, network connections, and the computer interface, based on configured policies.

Does local AI work entirely without internet? Not necessarily. A model can run on the computer while other steps use internet-connected services.

Are all the new features already available? No. MXC is generally available now, GitHub Copilot’s hybrid integration will have an experimental version in October, and new Copilot features will arrive in the coming months.