Microsoft opens the Copilot harness behind GitHub Copilot to developers in Foundry, in public preview

Microsoft has made the "harness" that runs its Copilot agents available to outside developers. In posts published on October 28 on the Official Microsoft Blog and the Microsoft Foundry blog, the company presented the Microsoft Copilot harness as the orchestration layer that sits around AI models: it supplies instructions and context, runs the agent loop in a sandbox, coordinates tool calls and applies controls so that an agent can carry out long, multi-step tasks.

What is being offered

According to the main announcement, signed by Foundry corporate vice president Tina Schuchman and GitHub CTO Vladimir Fedorov, the same harness already powers Copilot Studio, the GitHub Copilot app, the command-line tool and VS Code. Microsoft says it supports more than 20 frontier models as well as bring-your-own-key access, and can send each task to the most suitable model or combination of models, though the post does not list the models. It also says the runtime was recently rewritten in Rust to reduce session overhead and time to first token.

Developers can use the harness as a fully managed service in Microsoft Foundry or through the Copilot SDK. In Foundry Agent Service it is in public preview for prompt agents, which are defined declaratively. The preview bundles four elements: reusable skills, Foundry and Model Context Protocol tools with explicit authentication and approvals, file and code execution with built-in subagents for bounded tasks, and agent definitions that can be kept in source control. Agents can be created in the Foundry portal, in VS Code, or with the Azure Developer CLI.

Limits and pricing

The harness has to be chosen when an agent is created and cannot yet be switched on afterwards; existing agents are unaffected. Teams that want their own framework, loop or middleware are pointed to hosted agents, for which Microsoft also made long-running resilience generally available. The runtime is free until December 1, 2026. From then, sandbox compute costs $0.0994 per vCPU-hour and memory $0.0118 per GiB-hour, with model inference and other tools billed separately; prices may vary by region.

How it differs from Copilot Fusion

The announcement is separate from Copilot Fusion, the GitHub Copilot feature that picks and combines models for each prompt inside Copilot itself. The harness instead exposes Microsoft's underlying agent infrastructure so companies can build and run their own agents on it. Supported regions, models and preview limits are documented separately, and Microsoft points mainly to benchmarks published by GitHub to back its performance claims.