Windows is introducing three platform-level primitives to make AI agents safer: agent identity (agents run as a distinct security principal separate from the user), containment (via the Microsoft Execution Containers library, MXC, which dynamically composes sandboxes proportionate to the risk of each action), and manageability (IT admin policies, audit trails, and integration with Microsoft Defender, Entra, and Intune). The GitHub Copilot CLI and SDK are shipping sandbox support as an experimental feature, blocking dangerous file system and network actions from agents. The session includes live demos of a malicious agent hitting sandbox boundaries, a Hyperlight-based micro-VM container, and Copilot CLI sandbox commands. The core argument is that prompt engineering alone cannot govern autonomous agents — OS-level enforcement is required.