Windows becomes a system designed for AI agents, here's how

The PC of tomorrow runs agents, these AI which no longer just respond but which act, click, open files for you. Microsoft has also presented one, Scout, this assistant which acts on its own. But an annoying question lingers behind: an autonomous program which wanders around in your system is also a new gateway to trouble. Microsoft has just presented its parade.

It's called Microsoft Execution Containers, or MXC, now in preview. The idea is simple: instead of letting an AI agent roam freely in Windows, we lock it in a sandbox, a partitioned environment from which it cannot escape. The nuance that changes everything: this compartmentalization is imposed by the operating system itself. You describe the rules once, and Windows enforces them wherever your agents run. The agent may want to overflow, the wall is held by the OS, not by its good will.

There remains one point that deserves to be clearly stated: MXC is not a product that you buy, but a development kit and a rules model integrated into Windows and the Linux subsystem (WSL). Concretely, it offers several levels of partitioning, from simple process isolation — already used by the GitHub Copilot command line interface — to micro-virtual machines. Process and session isolation will arrive for Windows Insiders shortly after Build.

OpenShell from Nvidia on it, OpenClaw running inside

This plumbing is already in use. OpenClaw, an agent execution framework, relies on MXC to chain tasks within these Windows-managed boundaries. This is precisely where OpenShell, Nvidia's software layer, comes in, which adds rule management, routing of requests to the right model, and above all the masking of personal data. Simply put, OpenShell can erase your sensitive information from a request before it goes elsewhere. This is the building block that was missing to answer the question “but where does my data go when the agent is working?” ".

Microsoft is not alone in pushing MXC: the firm cites five partners already on board, including OpenAI, Nvidia, Manus, Nous Research (at the origin of the Hermes agent) and the open source project OpenClaw. Enough to give credibility to a technical brick which, on paper, could remain a dead letter without real adoption by developers.

The whole thing forms what Microsoft calls "native Windows for agents", and that gives meaning to the entire strategy that accompanies it: from the three ranges of Windows PCs presented as the biggest upheaval in 40 years to the seven in-house AI models that Microsoft has just unveiled to the seven in-house AI models that Microsoft has just unveiled A machine packed with memory to run large models locally is only worth it if you can entrust it with agents without trembling.

So, should we be reassured? In part, yes. This is exactly the kind of basic response we expected before releasing autonomous agents on our machines, and the approach, partitioning imposed by the OS rather than blind trust, is the right one. But we are talking about previews, and a sandbox is only as good as its real solidity against malicious uses. Microsoft has laid the right foundations. For once, we start with security rather than tinkering with it after the fact. That's already it.

To go further

Copilot on Windows 11: you will now have the last word before accessing your files