Microsoft announced on October 28, 2026, the launch of the Microsoft Copilot harness, making the orchestration layer originally built by GitHub available to developers and enterprise customers through Microsoft Foundry.
What the harness does
Microsoft describes the harness as the orchestration layer surrounding an AI model, providing instructions, context, execution environment, and controls for agentic work. The harness supports open standards including MCP, agent skills, and agent plugins.
Originally built by the GitHub engineering team, the harness powers Microsoft products such as Copilot Code, Copilot Studio, the GitHub Copilot app, the CLI, and VS Code.
Technical details
The harness is built to perform across model families, supporting over 20 frontier models alongside bring-your-own-key options. Microsoft states that it intelligently routes work to the right model based on the task.
Microsoft recently ported the harness runtime to Rust, which the company says has resulted in performance improvements and reductions in both session overhead and time-to-first token.
Availability in Foundry
Developers and enterprise customers can download the SDK or run the harness in the cloud, managed by Microsoft Foundry. Under this setup, Microsoft runs the agent loop and sandbox with built-in governance and optimization.
Microsoft notes that its platforms remain open and interoperable, allowing users to continue bringing in other harnesses and frameworks.
Key facts and where they come from
- Microsoft is making the GitHub-built Copilot harness available via Microsoft Foundry.
And today, we’re making that harness available through Microsoft Foundry.
- The harness supports over 20 frontier models plus bring your own key options.
The harness is built to perform across model families, with 20+ frontier models plus bring your own key
- Microsoft ported the harness runtime to Rust to improve performance and reduce overhead.
We recently ported the harness runtime to Rust, which has resulted in industry-leading performance and reductions in both session overhead and time-to-first token.
