Microsoft Opens the Agent Harness Behind GitHub Copilot to Foundry Developers, Free Until December 1
Microsoft introduced the Copilot harness on October 28, the agent loop it says runs Copilot Code, Copilot Studio, the GitHub Copilot app, CLI and VS Code, and put it into public preview in Foundry Agent Service. It supports more than 20 frontier models and bring-your-own keys. Runtime is free until December 1, when sandbox compute and memory billing begins.
Microsoft is selling the part of its coding agents that it argues matters as much as the model. On October 28 it introduced the Microsoft Copilot harness, the agent loop and execution layer it says powers Copilot Code, Copilot Studio, the GitHub Copilot app, the Copilot CLI and VS Code, and made it available in public preview to developers through Foundry Agent Service.
What the harness does#
In Microsoft's description, the harness coordinates model and tool calls, manages working context and files, and runs work in a sandbox that Microsoft operates. Developers define agents declaratively and bring their own skills, which package instructions and scripts, plus tools connected through Foundry or the Model Context Protocol, with configured authentication and approvals. Agents can delegate bounded tasks to built-in subagents.
Microsoft says the harness works with more than 20 frontier models and with customers' own keys, and that it rewrote the runtime in Rust. Microsoft's example skill is a SKILL.md file, the layout used by Agent Skills, the standard Anthropic donated to the Agentic AI Foundation last week; Anthropic counts GitHub Copilot and VS Code among the platforms that support it.
The multi-model pitch runs into a policy question: OpenAI's revised usage terms will require approval for model routers and AI gateways that give third parties direct model access, from November 20. Microsoft did not address that in its posts.
Terms of the preview#
Harness runtime is free until December 1, 2026. From then, sandbox compute is billed at $0.0994 per vCPU-hour and memory at $0.0118 per GiB-hour, on top of model inference and other tools, Microsoft says. During the preview, existing agents are unchanged and an agent cannot be switched to or from the harness after it is created.
Limits#
- Claims that the harness improves task completion are Microsoft's; it published no comparative benchmarks in the posts we read.
- Supported models and regions are listed in Microsoft's documentation, which we did not review.
- Whether OpenAI's November 20 policy applies to Foundry is not stated by either company.