Architecture overview

Important

Windows 365 for Agents is in public preview. The feature is under active development and might change before general availability.

The Windows 365 for Agents architecture provides a unified platform that supports two primary interaction models:

  • Human users who work interactively with Cloud PCs through a chat-based experience.
  • Agentic applications (such as Microsoft Copilot Studio (MCS), Project Opal, and Manus) that autonomously create, claim, and operate Cloud PCs on behalf of users or workflows.

IT administrators and agent makers configure and manage the underlying Cloud PC pool, while end users and agents access Cloud PCs on demand.

Core components

The platform is organized into four cooperating subsystems. Each subsystem owns a distinct stage of the Cloud PC for Agents lifecycle.

Computer-Create: provisioning

Computer-Create is responsible for creating and maintaining the Cloud PC agent pool. It's the control plane that IT admins and agent makers interact with.

Key elements:

  • Graph API, Microsoft Intune admin portal, Microsoft Admin Center: Administrative surfaces for configuration, policy, and chargeback.
  • Cloud PC pools: A collection of provisioned Cloud PCs for Agents.
  • Enterprise device provisioning: The same process used for Windows 365 Enterprise provisioning handles Microsoft Entra and Intune enrollment for each Cloud PC for Agents.
  • Scalable, Hosted on behalf of: The infrastructure fabric that provisions compute cost-efficiently at scale.
  • Virtual machines (Windows): The end workloads. Each VM runs an on-box CUA client that enables agentic control.

Computer-Get: assignment

Computer-Get brokers available Cloud PCs from the pool to the caller that needs one.

Key elements:

  • MCP server: Exposes Cloud PC acquisition capabilities through the Model Context Protocol (MCP), making them callable by agents.
  • Check-out / check-in: Reserves a Cloud PC for a session and returns it to the pool when the session ends.
  • Agentic cloud assignment: Matches requests to the optimal Cloud PC based on capability, region, and availability.

Computer-Do: actions

Computer-Do executes commands on an assigned Cloud PC. It's the plane through which agents drive the operating system.

Key elements:

  • MCP server: Exposes the action API (click, type, navigate, run) to orchestrators.
  • Relay and protocol: Transports action requests from the agent to the on-box CUA client running inside the target Cloud PC.

Computer-See & Computer-Take Control: access and control

Computer-See & Computer-Take Control delivers the interactive pixel and device experience to humans.

Key elements:

  • IC3 media: Real-time media stack for audio, video, and peripheral redirection.

Entry points

  • Chat UX: The human-facing entry point. A user converses with the system and is connected to a live Cloud PC session through Computer-See & Computer-Take Control.
  • Agentic app: A host that contains a model and an orchestrator. The orchestrator calls Computer-Get to claim a Cloud PC and Computer-Do to operate it.
  • IT admin and agent makers: Administrative entry point into Computer-Create for pool configuration and lifecycle management.

Next steps