Microsoft introduces the Copilot harness for multi-model agent work

Microsoft announced on October 28 that it's making its Copilot harness, the orchestration layer that sits around AI models, available to developers through Microsoft Foundry. Microsoft describes a harness as everything that surrounds a model: the instructions, context, execution environment, and controls needed for an agent to actually get work done, including support for open standards like MCP and agent skills.

The harness was originally built by GitHub's engineering team and already powers Copilot Code, Copilot Studio, the GitHub Copilot app, the CLI, and VS Code; Microsoft says it has now adopted it company-wide. It supports more than 20 frontier models plus the option to bring your own API key, and can route tasks to whichever model fits best rather than locking developers into one. Microsoft says it recently ported the harness runtime to Rust, which it says delivered industry-leading performance along with cuts to session overhead and time-to-first-token.

Through Foundry, developers and enterprise customers can now build agents on this same harness in a fully managed way, or download the SDK to run it themselves, without having to build and maintain their own orchestration layer as new models arrive. Microsoft says the platform stays open, so developers can still bring their own harness or framework instead.

Why it matters

Giving outside developers the same orchestration layer that runs GitHub Copilot removes a major engineering burden for anyone building multi-model agents.

Source: Microsoft