Microsoft has announced wider access to the software system that coordinates its Copilot agents, offering it through Microsoft Foundry and an SDK. The October 28 announcement calls this system the Copilot harness. It is the surrounding machinery that gives a model instructions, context, tools and an execution environment so it can carry out a task.
The company says the harness originated with GitHub's engineering team and now supports products including Copilot Code, Copilot Studio, GitHub Copilot, its command-line interface and Visual Studio Code. Microsoft presents the shared system as a way for developers to build agents without maintaining all of that coordinating software themselves.
Microsoft says the runtime supports more than 20 frontier models, alongside a bring-your-own-key option. It also describes support for MCP, agent skills and plugins. Those are mechanisms for connecting agents to tools and reusable instructions. Model choice and the surrounding application remain separate decisions even when they share an execution system.
Developers can download the SDK or use a cloud deployment managed through Foundry, according to the announcement. Microsoft says it handles the agent loop and sandbox while customers supply instructions, skills, tools and model choices. It also says it recently moved the runtime to Rust and reports performance improvements, without an independent evaluation in this article.
For a development team, the practical question is how much operational work the managed system actually removes for its own application. An agent still needs suitable permissions, clear success criteria and a way to detect mistakes. Reusing a runtime does not establish that every tool action or generated result is appropriate for a particular business.
The announcement does not provide a universal price or guarantee that one configuration will be cheapest. Teams comparing approaches would need to measure their own task completion, response time and usage costs. The significant change is access to Microsoft's shared coordination layer; the quality of a finished agent still depends on the system built around it.