Microsoft brings its GitHub-built Copilot harness to Foundry

Microsoft describes a “harness” as everything surrounding an AI model: the orchestration layer that supplies an agent with instructions, context, a runtime environment, and the controls needed to get real work done, alongside support for open standards like the Model Context Protocol and agent skills. Competing effectively now requires working across multiple models, the company argues, but building a harness that performs consistently across different models and environments is difficult — it has to balance performance, quality, and cost, and then be maintained as the field keeps moving, since a leading model today can fall behind tomorrow. By exposing the Copilot harness through Foundry, Microsoft says developers get a fully managed way to build agents that reason, use tools, and complete long, complex tasks, picking whichever model fits a given task without swapping out the harness or the rest of their stack.

The harness was designed from the start to be fast, token-efficient, and predictable, and Microsoft says it has delivered the highest satisfaction scores among GitHub’s own developer users — which is why the company is now standardizing its Copilot products on the same foundation. It supports more than 20 leading models plus the option to bring your own API key, and Microsoft says it will keep directing work intelligently to the right model or models for each task rather than leaving developers to track and evaluate new models themselves. The company recently migrated the harness’s runtime to Rust, which it says delivered industry-leading performance and cut both per-session overhead and time-to-first-token, and because the Copilot harness already underpins multiple Microsoft products, Microsoft says it has been proven at the scale of millions of users and hundreds of real-world use cases.

Developers can download the SDK directly or run the harness in the cloud with everything managed through Foundry, and Microsoft says its platforms remain open and interoperable, so developers can still use any other harness or framework they prefer. With the Copilot harness, Microsoft runs the agent loop and sandbox with governance and optimization built in from day one, while developers supply the instructions, skills, tools, and model choice.

Why it matters

By opening up the same harness that runs its own Copilot products, Microsoft is betting that shared, continuously improving infrastructure — rather than a single model — is what will differentiate agent platforms as model choice keeps expanding.