Microsoft has made Microsoft Execution Containers (MXC) generally available for organizations running local AI agents on Windows. The problem is the authority those agents can inherit: an agent granted broad endpoint privileges can reach files, network destinations, processes and user interactions beyond its assigned task. MXC is Microsoft’s operating-system-enforced way to put limits around that work.
For endpoint teams, the immediate question is which local agents run on their Windows devices and what each task actually needs to reach. Microsoft’s Windows AI documentation describes a platform for on-device models and agent integration; those capabilities make the agent’s local access worth scoping separately from its instructions.
An OS limit around agent work
Microsoft describes MXC as a least-privilege containment layer across files, networks, processes and user interactions. Its roundup says those boundaries take the place of trusting an agent to stay within its intended limits. The control sits at the operating-system layer, where access can be restricted even when the agent asks for something outside the task. Microsoft also describes additional isolation options for higher-risk scenarios.
An agent may be instructed to summarize a local document, but the instruction alone is a poor access policy for the endpoint hosting it. Decide which document locations and network destinations its work requires, then configure the container policy for those requirements. Apply the same reasoning to process access and actions involving the user. A broad allowlist gives the agent the reach you were trying to avoid.
Pilot policies before fleet enforcement
Microsoft says organizations can tune and validate MXC policies before enforcing them broadly. Start with one known Windows agent and a real, approved task. Record the resources that task needs, make those resources available through the policy, and restrict unrelated access. In the pilot, verify the task still finishes. Then run an authorized access test against a file or destination outside its assigned work and confirm the policy blocks it. Treat this as your own deployment acceptance test, not a vendor-published test result.
Microsoft describes integration with Intune, Entra and Agent 365 as upcoming. Plan your current MXC pilot around the controls you can configure and inspect today; avoid treating those future integrations as deployed fleet governance.
Keep each agent inside its task
Inventory the Windows local agents you intend to run and identify their legitimate file, network, process and user-interaction needs. Configure MXC for a representative task, then retain the policy and the pilot’s allowed-task and denied-access results as the control record. That gives endpoint owners a concrete way to compare intended authority with observed behavior before wider enforcement. Agent discovery and container policy answer different questions: finding an agent tells you where to start; a tested policy tells you what that agent could reach in the pilot.
A container is useful only when its limits match the work. Keep the approved task working, and require the out-of-scope access test to fail before extending that policy to more devices.