Developers can now build agents on the same machinery that runs Microsoft’s own Copilot products. Microsoft said on 28 October that the Microsoft Copilot harness, the layer behind Copilot Code, Copilot Studio, the GitHub Copilot app, the Copilot CLI and VS Code, is available through Microsoft Foundry, either as a downloadable SDK or as a managed service in the cloud.
The move turns an internal piece of plumbing into a product. A harness is everything wrapped around an AI model to make it an agent: the instructions, the context it carries, the tools it can call, the sandbox it runs in and the loop that keeps it working. Microsoft is betting that companies would rather rent the one it already maintains than build and keep up their own, especially as the best model for a job keeps changing.
A GitHub-built layer now runs across Microsoft
The announcement, written by Tina Schuchman, corporate vice-president for Foundry, and Vladimir Fedorov, GitHub’s chief technology officer, says GitHub engineers built the harness and Microsoft has “adopted company-wide”. It came on the same day GitHub made Copilot Fusion generally available, its system for routing coding tasks between models. The Microsoft post says that kind of multi-model orchestration was only possible because the company built its own harness.
Microsoft says the harness works with “20+ frontier models plus bring your own key”, so customers can also use their own model accounts, and that it can route work to the right model or models for a task. It supports the Model Context Protocol (MCP), agent skills and plugins. Microsoft also says it recently ported the harness runtime to Rust, cutting session overhead and the time to the first token, though it gave no figures.
Satya Nadella, Microsoft’s chief executive, described it on X as “a managed service for building long-running agents that reason, use tools, and get work done”.
Free in preview, billed from 1 December
The Foundry Blog post gives the practical details. The harness is in public preview in Foundry Agent Service, where developers define an agent declaratively and Microsoft runs the agent loop and sandbox. It brings four building blocks: reusable skills, connected Foundry and MCP tools with explicitly configured authentication and approvals, the ability to delegate bounded tasks to built-in subagents, and agent definitions that can sit in source control alongside an application’s code.
There are limits. The harness has to be chosen when an agent is created; existing agents are unchanged and cannot yet be switched over, and Foundry controls the sandbox’s CPU, memory and idle settings.
Harness runtime is free until 1 December 2026, when consumption billing starts. From then, sandbox compute costs $0.0994 per vCPU-hour and memory $0.0118 per GiB-hour, with prices varying by region and agreement. Model inference and other tools are billed separately throughout. Microsoft also made long-running resilience for hosted agents, which restarts interrupted work, generally available.
One early customer, Chris Evans, chief technology officer of Confluence Technologies, said his team “had a solution running in less than a week”.
Renting the harness means renting Microsoft’s choices too
The pitch rests on a real problem. Teams building agents tend to rebuild the same scaffolding, and a frontier model that leads today may trail next month. A harness that already handles many models lets a team swap the model without rewriting the rest, and Microsoft carries the maintenance.
The trade-off is control. Microsoft’s Foundry Blog calls the harness its “opinionated approach” to complex work, and the decisions it makes about context, tool calls and when to escalate between models are decisions the customer no longer owns. Automatic routing also makes costs harder to predict, because the model behind a task is chosen at run time. The free preview hides that until December, so teams should measure model spending, not just harness fees, before relying on it.
There is also an open question about whether elaborate harnesses help at all. A recent academic preprint found that a minimal coding agent matched far more elaborate harnesses on its test set, pointing to the model as the main driver. Microsoft’s performance claims for its harness are its own, with no published figures. For companies that already worry about agents behaving like insider risks, the per-tool approvals are the part to examine first.
A crowded market for agent plumbing
Microsoft is not alone in selling the layer around the model. Cursor recently released its Agent SDK, which turns agents into testable code projects, and GitHub’s Fusion pushes the same multi-model idea inside Copilot itself. Microsoft says its platforms stay open, so customers can still bring another harness or framework.
The date that matters now is 1 December, when the free preview period for harness runtime ends and billing begins.
Sources
- Official Microsoft Blog, “Introducing the Microsoft Copilot harness”, Tina Schuchman and Vladimir Fedorov, 28 October 2026
- Microsoft Foundry Blog, “Introducing the Microsoft Copilot harness in Foundry Agent Service”, 28 October 2026
- Microsoft Developer on X, 28 October 2026
- Satya Nadella on X, 28 October 2026
Related stories
- GitHub Makes Copilot Fusion Generally Available and Previews Projects
- Cursor’s Agent SDK Turns AI Agents Into Code Projects You Can Test
- Do AI Agents Need Elaborate Harnesses? A Minimal-Agent Study and Two New Toolkits
- Microsoft Treats AI Agents as Insider Risks, Adds Guardian Agents to Watch Them
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
