Microsoft’s Copilot Harness Makes Agents Easy. Switching Isn’t Supported Yet

Building an AI agent just got easier. Moving it later might not be.

On October 28, 2026, Microsoft opened the GitHub Copilot agent harness to outside developers. The harness is the engine behind Microsoft’s own Copilot tools. Now any team can use it through Microsoft Foundry, as Microsoft’s Foundry blog explains.

The launch is real and useful. But the fine print matters more than the headline. It says something important about how you will pay for this convenience. Not in dollars, but in options.

What Microsoft is actually offering

A model can reason about a task. Finishing the task takes more. The agent needs context, tools, a place to run code, and rules. A “harness” is the layer that supplies all of that. According to The New Stack, Microsoft’s harness manages model and tool calls, tracks context, and keeps agents within configured policies.

Microsoft now runs that layer for you. In its own words, it runs the agent loop and the sandbox. You bring the knowledge, the procedures and the tools. The service is called the Microsoft Copilot harness in Foundry Agent Service.

The public preview has four building blocks:

• Reusable skills that package instructions, scripts and supporting files.
• Connected tools, including Model Context Protocol (MCP) tools, with authentication and approvals you configure.
• Execution and delegation, so agents can work with files and code and hand tasks to built-in subagents.
• Source-controlled agent definitions, so the setup can live next to your application code.

Microsoft’s walkthrough shows how little code this takes. A Markdown file describes a research skill. A short azure.yaml file names the model, tools and instructions. Then you pick the harness type and run one deploy command.

Early customers sound pleased. Confluence Technologies’ CTO, Chris Evans, said his team had a solution running in less than a week, as Unite.AI reports from Microsoft’s post. Keep in mind that these are vendor-selected testimonials.

Preview status and the price tag

This is a public preview, not a finished product. Microsoft’s own documentation says previews come without a service-level agreement. It also says Microsoft doesn’t recommend them for production workloads.

Money is simple for now. Harness runtime is free before December 1, 2026. Model inference and separately billed tools and services still cost money. From December 1, sandbox compute is billed at $0.0994 per vCPU-hour and memory at $0.0118 per GiB-hour. Microsoft says prices may vary by region and agreement.

Notice who decides how big that sandbox is. Microsoft does. That detail leads straight to the real story.

The catch: you choose once

When you create an agent, you select the Copilot harness. After that, the choice is fixed. The docs say you can’t switch an existing agent during the initial public preview. Creating a new version of the agent doesn’t change it either.

Microsoft is open about this. The Foundry blog says:

“switching the harness after creation isn’t supported yet” — Jeff Hollan, Viswajeeet Balaji and Anthony Chu, Microsoft Foundry blog

The word “yet” is doing a lot of work. It suggests Microsoft may fix this. It has not promised a date, and none of the pages we read gave one.

Two more limits stand out. Foundry manages the agent’s CPU, memory and idle settings. And the docs say bring-your-own-model isn’t currently supported with the Copilot harness. You pick from compatible models deployed through Foundry.

That is why The New Stack’s headline lands. Agents are easy until you want to switch.

Two kinds of lock-in, one of them small

Lock-in is not one thing here. We see two layers, and they carry very different risk.

The first layer is your work: instructions, skills, tool definitions. These are Markdown and YAML files that can sit in Git. The docs say the agent stays a prompt agent that you call through the Responses endpoint. Unite.AI reports that Microsoft supports open standards such as MCP and agent skills. This layer looks fairly portable. We have not tested moving it to another runtime, so treat that as our reading, not a guarantee.

The second layer is the runtime: the agent loop, the sandbox, the resource limits and the workspace state. You don’t own any of it. The docs add that the sandbox is not a durable output store, so important files must live somewhere else.

Here is our view. The risk sits in behavior, not syntax. A skill written for one loop may act differently in another. You can copy the files. You can’t copy the engine. So the cost of leaving is mostly testing and re-tuning, and that cost grows with every agent you build.

Microsoft’s escape hatches

Microsoft offers a way out of the managed harness, and it stays inside Foundry. Foundry Agent Service has three approaches. Plain prompt agents work when instructions and supported tools are enough. Prompt agents with the Copilot harness add skills, files and subagents. Hosted agents let you bring your own framework, agent loop, middleware and dependencies while Foundry handles hosting.

Hosted agents also got an upgrade. Microsoft announced general availability of long-running resilience for hosted agents. Foundry restarts interrupted work, but your code must save progress.

There may be a second exit. Unite.AI reports that Microsoft says the Copilot SDK is available to download and that its platforms stay open to other harnesses and frameworks. Unite.AI also reports support for more than 20 frontier models plus bring-your-own-key. That differs from the Foundry docs, which say bring-your-own-model is not currently supported in the managed preview. These claims may describe different products, so check which one you are using.

So the choice is a ladder of control. More managed means less work now and fewer exits later. If you think you will need your own loop in six months, start with a hosted agent. If a standard research or coding-style agent is enough, the harness will save you real time.

What to do before December 1

You don’t need to avoid the harness. You need to use it with an exit plan. We suggest five steps:

• Keep skills, instructions and agent definitions in Git, not only in the portal.
• Write outputs to durable storage. Never treat the sandbox as the archive.
• Keep production workloads off the preview, as Microsoft’s own docs advise.
• Build a throwaway agent now, while runtime is free, and measure what it really costs.
• Decide per agent: harness or hosted. Write the reason down.

If you are new to this, start small. Our guide on why your first AI agent will fail explains why. Sandboxing is also a theme we covered in Your AI Agent Needs a Cage. Microsoft Just Built One.

Now watch four things. First, whether Microsoft removes the “yet” and lets teams switch harnesses. Second, whether bring-your-own-model arrives in the managed service. Third, when the product leaves preview. Fourth, what the first real bills look like after December 1.

Microsoft has made agents cheaper to start. The open question is whether they stay cheap to change. Until switching works, treat your harness choice as architecture, not a setting.

Sources & Credits

Text credit & news reporting source: Microsoft Foundry Blog / Microsoft Learn / The New Stack / Unite.AI.

• Microsoft Foundry Blog: Introducing the Microsoft Copilot harness in Foundry Agent Service
• Microsoft Learn: Use the Microsoft Copilot harness with prompt agents – Microsoft Foundry
• The New Stack: Microsoft’s new Copilot harness makes agents easy — until you want to switch
• Unite.AI: Microsoft Makes Copilot Harness Available Through Foundry

Enjoyed this breakdown? Stay ahead of the curve with our technical analyses and news. Join our community by subscribing to the CVisiona Newsletter for weekly deep dives, and subscribe to our CloudMelon Vis YouTube Channel for video breakdowns of Cloud-Native, AI, and Tech Enterprise news.

Question for you: If you build agents on a managed harness today, how would you plan to leave it later: rebuild, re-test, or never leave? Tell us in the comments.