Microsoft opened the Copilot Harness to outside developers on October 28, and this one is worth your attention. The harness is the agent orchestration runtime that has been quietly running GitHub Copilot, Copilot Studio, VS Code, and the CLI at Microsoft scale — routing billions of tokens daily through multi-model routing, tool dispatch, sandbox execution, and subagent coordination. It is now accessible to any developer via Foundry Agent Service, and the window to try it without incurring sandbox compute costs closes December 1.
What the Harness Actually Is
The Copilot Harness is not a new API wrapper or a rebranded Azure service. It is the entire agent loop: the layer that decides which model runs which task, dispatches tools, manages the sandbox environment, and coordinates handoffs between subagents. GitHub engineers rebuilt it from TypeScript to 800,000-plus lines of Rust over 14.5 weeks, using Copilot agents to do the porting work. Session round-trip latency dropped from 5.25 seconds to 292 milliseconds in-process. The performance numbers are real because the scale is real.
That backstory matters for evaluation: you are not adopting an experiment. You are getting access to infrastructure Microsoft has been stress-testing against its own product lines for over a year.
What Developers Get
The harness exposes four components:
- Skills — Reusable Markdown files with YAML frontmatter stored in a
skills/directory. These package instructions, scripts, and resources into named procedures the agent can invoke across tasks. Skills are source-controlled and portable. - Connected Tools — Foundry-native tools plus full Model Context Protocol (MCP) integration, giving agents access to more than 10,000 MCP servers without custom auth plumbing.
- Built-in Execution — Five harness tools baked in:
filesystem_read,filesystem_write,shell,web, andsubagents. Subagents let a coordinator agent delegate specific tasks to child agents and collect results. - Declarative YAML Agents — The entire agent definition — model choice, skills list, tool connections, instructions — lives in a versioned YAML file deployable with
azd deploy --all.
Getting started takes three paths: the Foundry portal for a GUI onboarding, the Azure Developer CLI for code-first teams, or the Foundry Toolkit extension in VS Code. Once deployed, azd ai agent invoke runs the agent.
The Lock-In You Should Know About
The New Stack’s headline put it plainly: “makes agents easy — until you want to switch.” That is accurate.
When you create a prompt agent in Foundry and select the Copilot Harness, that choice is permanent. You cannot migrate to a different framework in a later version. Microsoft also controls the sandbox: CPU limits, memory caps, and idle timeouts are not developer-configurable in the managed offering. Model selection is restricted to what Foundry supports; bring-your-own-model is not available with the Copilot Harness path. Regional availability is narrower than Foundry Agent Service broadly.
The escape hatch is real but requires deliberate planning. Foundry’s hosted agents option lets teams run their own frameworks — LangChain, CrewAI, custom — with Microsoft handling the hosting. The harness is opt-in, not mandatory. If you choose it, go in knowing you are making a runtime commitment.
MCP Skills Are Portable
There is a meaningful counterpoint to the lock-in concern. The logic layer — skills (Markdown files), tool connections (MCP servers), and agent definitions (YAML) — is extractable and readable outside the harness runtime. If MCP stays the dominant agent tool protocol, skills built for the Copilot Harness will carry over to other runtimes. The proprietary piece is the execution environment itself, not the agent’s knowledge or capabilities.
That is not a guarantee, but it is a better story than full black-box lock-in.
Pricing and Timeline
The sandbox compute charge does not kick in until December 1, 2026. After that, billing runs at $0.0994 per vCPU-hour and $0.0118 per GiB-hour. Model inference has always been billed separately and continues to be. There is no SLA for the preview, and Microsoft explicitly flags it as not ready for production workloads.
That last point is the most important one to act on: the four weeks until December 1 are designed for experimentation, not shipping. Use the window to evaluate whether the harness fits your agent architecture before the cost meter starts and before you commit to the framework permanently.

