Microsoft is opening the control software behind its own Copilot products to outside developers. The Official Microsoft Blog announced on 28 October 2026 that the Microsoft Copilot harness is now available through Microsoft Foundry, the company’s platform for building and running AI applications. The post was written by Tina Schuchman, who leads Foundry, and Vladimir Fedorov, GitHub’s chief technology officer.
A harness is the software wrapped around an AI model. The model supplies the thinking, and the harness does everything else: it hands the model its instructions, decides what context the model sees, connects it to tools, runs the code it writes inside a safe sandbox, and keeps the job moving across dozens of steps. Without one, a model answers a question and stops. With one, it can chew through a long task, like fixing a bug that touches a dozen files.
The harness started inside GitHub. Microsoft says it now runs underneath Copilot Code and Copilot Studio, as well as the GitHub Copilot app, the command-line tool, and VS Code. Foundry customers can download an SDK or run it in the cloud, where Microsoft operates the agent loop and the sandbox. The customer brings instructions, skills, tools, and a choice of model.
That choice is wide on paper. The post says the harness works with more than 20 frontier models, plus a customer’s own API keys, and can send each task to whichever model suits it. It also supports open standards such as MCP, the protocol for plugging tools into AI agents, along with agent skills and plugins.
The pitch is mostly about upkeep. Models leapfrog each other every few months, and a team that wrote its own harness has to retest and retune it each time a new one lands. Microsoft is offering to carry that work and to improve the harness continuously, so that every product built on it benefits at once. The authors add that the runtime was rewritten in Rust, which cut session overhead and the wait before the first token appears.
Almost everything in the post is Microsoft describing its own product. It calls the Rust result “industry-leading performance” and publishes no comparison and no figures. It says the harness has been tested on millions of users, without saying how many of those are paying Foundry customers. It also does not give a price for the hosted version.
The post only hints at the market angle. Model makers compete on who has the smartest model, and Microsoft is now selling the layer that makes any of them useful for long work. A company that builds its agents on a Microsoft-run harness can swap models freely, but its agents and their governance settings live on Microsoft’s infrastructure. The post says the platforms stay open and customers can bring any other harness they like, which is the right promise to make and a hard one to test before someone tries to move.
Until Microsoft prints a price for the hosted version, the build-or-rent math for any team with its own agent loop has one blank cell.
Reported by the Official Microsoft Blog on 28 October 2026.
