Microsoft Tests 'Scout,' a Hybrid AI Agent That Runs Directly on Your PC

Microsoft is testing a new kind of AI assistant called Scout, which marks a significant departure from purely cloud-based tools. Scout is designed as a persistent agent that runs directly on a user's desktop, a move that could enable more autonomous, background-based tasks.

This local architecture gives it direct access to the file system and shell, allowing it to perform operations that cloud APIs can't, while still connecting to the user's Microsoft 365 environment. The design introduces new considerations for corporate governance, resource management, and especially cost.

Scout operates as a desktop client for Windows and macOS and is built on the open-source OpenClaw agent framework. It's provisioned with permission to read and write local files and can execute PowerShell commands or run self-generated Python code. This local capability is paired with API access to a user's Microsoft 365 data, including Teams, Outlook, and OneDrive. A memory layer called Work IQ provides context, allowing the agent to learn user patterns and project priorities over time. Users can also extend its functionality by defining custom skills in markdown files.

Task execution is triggered in two ways. Automations function like scheduled jobs, running a defined workflow based on a set time or a file system event. Heartbeats are simpler, acting as a recurring poll that executes a single prompt on a fixed interval for continuous monitoring, such as triaging an inbox every hour.

Governance, Deployment, and Cost Structure

The security model is centered on identity. Each Scout agent is provisioned with a unique Microsoft Entra identity, making every action it takes, from API calls to local script execution, attributable to a specific, auditable service principal within the organization.

Administrators get a multi-layered control plane. They first use Microsoft Intune to configure and enforce boundaries on target devices. The agent then operates within existing Microsoft Purview data protection policies, respecting sensitivity labels and data loss prevention (DLP) rules. High-privilege actions, such as running a script that modifies local files, require a user-facing prompt for explicit approval, creating a human checkpoint.

Deployment is currently limited to an experimental release for organizations in Microsoft's "Frontier" program. The prerequisites for an organization include enrollment in the program, configuration of a specific Intune policy, and submission of a data processing attestation. Users need a Windows 11 or macOS device, a Microsoft 365 account, and a GitHub Copilot license.

The billing model is a key detail: it is not part of the standard Microsoft 365 Copilot subscription. Instead, Scout operates on a consumption basis, drawing credits from a user's existing GitHub Copilot Business or Enterprise license pool. This means costs scale directly with the agent's activity. High-frequency Heartbeats and complex Automations will consume credits more quickly, requiring organizations to monitor the agent's autonomous workload to manage costs.