Microsoft Turns Windows into a Platform for Local AI Agents

Microsoft Intensifies Local AI Integration on Windows

At the Build 2026 conference, Microsoft continued to advance its project for integrating local artificial intelligence into Windows. After establishing models, neural processing units (NPU), and developer tools, the company now aims to better manage the execution of AI agents directly on machines. This initiative began at Build 2025, where Microsoft introduced elements like Windows AI Foundry, Foundry Local, Windows ML, and support for MCP to connect agents to local applications.

The Ignite event marked another milestone with the introduction of Agent Workspace, a dedicated space in Windows where agents can access applications and files under a separate account from the user. At Build 2026, Microsoft refocused its efforts on the local execution of AI agents and the role that Windows can play when these agents operate directly on the computer.

Microsoft Execution Containers: A New Era for AI Agents

After laying the groundwork for local AI with models and developer tools, Microsoft is tackling the next step: enabling agents to act directly on the computer. The main innovation is the Microsoft Execution Containers (MXC) technology, currently available in preview. This technology introduces a policy layer in Windows to define the confinement rules for AI agents, which the system then enforces.

The idea is to prevent these agents from operating in poorly managed user sessions when executing code, accessing files, or interacting with the machine's network. Developers and IT administrators can define the execution conditions once, and Windows takes care of enforcing them through its native mechanisms. This aims to reduce the adjustments needed to run agents locally while maintaining strict control over identity, isolation, and governance.

MXC is expected to support agent runtimes like OpenClaw on Windows, which is now available in alpha on GitHub. The goal is to allow these agents to execute local multi-step task sequences within limits imposed by the OS, rather than in simple user sessions. NVIDIA is also collaborating with Microsoft on OpenShell, an open-source runtime for autonomous agents built on MXC to facilitate their execution in sandboxed environments. This runtime will also support more enterprise-friendly features, such as policy management, inference routing, and obfuscation of personally identifiable information.

Microsoft is also applying this confinement model to agents hosted in Foundry Agent Service. Each agent session has its own sandbox, persistent memory, and the ability to scale as needed. Whether the agent operates on Windows or in the cloud, the goal is to provide it with an isolated execution environment, predefined and tailored to its actions.

A New Development Surface for Running Large Models Locally

To support this software shift, Microsoft is also focusing on hardware. The Surface RTX Spark Dev Box is a development machine designed around the new NVIDIA RTX Spark, intended for sustained AI workloads such as long processing tasks, agent operations, or local fine-tuning of models.

Microsoft announces up to 1 petaFLOP of computing power and 128 GB of unified memory. According to the publisher, this configuration should allow for the local execution of models with up to 120 billion parameters, without relying on a cloud GPU instance. These figures should be taken as manufacturer promises, but they provide a clear idea of the target audience. This is not a consumer PC enhanced with AI, but a compact workstation for developers, AI teams, and organizations looking to increase local trials before moving heavier workloads to Azure.

The Surface RTX Spark Dev Box will be available starting this fall in the United States through the Microsoft website. It will come with WSL 2 preconfigured, native CUDA support, and direct GPU access from the Linux environment, as well as Visual Studio Code, GitHub Copilot, and several pre-installed developer tools.