Microsoft turns its Copilot agent loop into a managed service

Microsoft has launched the Copilot harness in Foundry Agent Service, turning the execution layer used across its Copilot products into managed infrastructure for other developers.

A model alone cannot complete complex work reliably: an agent also needs tools, files, working context, a sandbox and rules for continuing through many steps. Microsoft says the new public preview handles that loop, allowing teams to define an agent without rebuilding orchestration and execution from scratch.

Customers can package repeatable procedures as SKILL.md files, connect Foundry and Model Context Protocol tools, and let the main agent delegate bounded jobs to built-in subagents. Agent definitions can live in source control, while Microsoft operates the sandbox and maintains context across model and tool calls. Developers can choose a supported Foundry model without rewriting the orchestration layer.

The managed harness is aimed at workflows such as research, data analysis and internal operations that need files, code execution and multiple tool calls. Teams that need full control over their framework or runtime can still use Foundry’s hosted-agents option instead.

Microsoft also made long-running resilience generally available for hosted agents. The platform can restart interrupted jobs with their original input, while the customer’s code is responsible for saving progress and avoiding duplicate actions.

Harness runtime is free during the preview, with model inference and connected services billed separately. Microsoft plans consumption pricing for sandbox CPU and memory after the free period. The release gives enterprises a way to rent a complete agent loop rather than build and operate one themselves.