Microsoft Copilot Harness in Foundry Agent Service: Public Preview Facts, Pricing From December 1, and What Is Not Said

On October 28 Microsoft announced the Microsoft Copilot harness in Foundry Agent Service, in public preview. The pitch: Microsoft runs the agent loop and the sandbox, and you bring instructions, skills, tools and a model. The part most worth reading is the pricing note. Per Microsoft's Foundry blog, harness runtime is free before December 1, 2026, and sandbox compute and memory are billed after that.

I read three primary posts from the same day: the Official Microsoft Blog announcement, the Microsoft Foundry Blog post, and the GitHub changelog entry about the SDK. I did not read the Foundry docs pages on supported regions, models, tools and preview limits, and I did not deploy anything. This is what Microsoft says, not a test.

What Microsoft announced

The Official Microsoft Blog post, by Tina Schuchman and Vladimir Fedorov, says the harness was built by GitHub engineers and adopted company-wide. It says the harness powers Copilot Code, Copilot Studio, the GitHub Copilot app, the CLI, VS Code and more, and that it is now available through Microsoft Foundry.

The post defines a harness as everything around the model: the orchestration layer that gives an agent instructions, context, execution environment and controls. It says the harness supports open standards such as MCP and agent skills, as well as agent plugins.

Claims in that post that are Microsoft's own, which I cannot verify:

• It supports "20+ frontier models plus bring your own key", with routing of work to the right model.
• The harness runtime was ported to Rust, which "has resulted in industry-leading performance" and reductions in session overhead and time-to-first token. No numbers are given.
• It is "battle-tested" at the scale of millions of users.

The post also says the SDK can be downloaded or the harness run in the cloud, managed by Foundry, and that the platforms stay open so you can bring other harnesses and frameworks.

What the Foundry post adds

The Foundry Blog post is the more specific one. It says the harness is in public preview in Foundry Agent Service and lists four building blocks:

1. Reusable skills, packaging instructions, scripts and supporting resources.

2. Connected tools, meaning supported Foundry and MCP tools with explicitly configured authentication and approvals.

3. Execution and delegation: work with files and code, and delegate bounded tasks to built-in subagents.

4. Source-controlled agent definitions, created in Foundry or managed alongside application code.

It shows an example using the preview Azure Developer CLI surface. A SKILL.md file holds a research procedure, an azure.yaml declares the skill and an agent with harness: type: microsoft_copilot_preview, and azd deploy --all followed by azd ai agent invoke runs it. The example agent uses a web search tool and a code interpreter. The post notes you can also start in the Foundry portal or the Foundry Toolkit for Visual Studio Code.

It compares three approaches: prompt agents without the harness (instructions and supported tools are enough), prompt agents with the harness (skills, managed execution, files, subagents), and hosted agents (you bring or change the framework, loop, middleware, dependencies or runtime).

Hosted agents: long-running resilience

The same post says long-running resilience for hosted agents is generally available "this week". Its description: Foundry detects interruptions and restarts the work with its original input. Your code saves progress and intermediate results to pick up where it left off. If the agent updates another system, your workflow must prevent duplicate actions on restart. The post's example is checking whether a report already exists before creating it again. It also says conversation history alone is not enough to restore work in progress.

That is a useful contract to read plainly: Foundry restarts, you checkpoint. It is not automatic state recovery.

Pricing and preview limits

Per the Foundry post:

• Billing for the Copilot harness begins December 1, 2026, with consumption-based pricing.
• Harness runtime usage is free before December 1. Model inference and separately billed tools and connected services remain subject to their charges.
• From December 1, sandbox compute and memory are billed at USD $0.0994 per vCPU-hour and $0.0118 per GiB-hour. Prices may vary by region and agreement.
• During the preview you select the harness when creating the agent. Existing agents are unchanged, switching the harness after creation is not supported yet, and Foundry manages the execution environment's CPU, memory and idle settings.

The post does not say how many vCPUs or how much memory a typical session uses, so you cannot estimate a monthly bill from these rates alone.

The SDK rename

The GitHub changelog says GitHub Copilot SDK is becoming Microsoft Copilot SDK, the programmable interface for the Microsoft Copilot harness. The migration happens in phases, package names, namespaces and some documentation may keep the GitHub name during the transition, no immediate changes are required for existing users, and support stays the same.

What I would check

These are my untested suggestions.

1. Read the supported regions, models, tools and preview limits page the Foundry post links to, before you plan around it.

2. Run one small agent before December 1 and look at what the sandbox actually consumes, since billing starts then.

3. Decide who owns checkpointing if you use hosted agents. Write down what state you save and how a restart avoids duplicate writes.

4. Check your approvals setup for MCP tools. The post says authentication and approvals are configured explicitly. Confirm that matches what you expect.

5. If you use the GitHub Copilot SDK, watch the repository and package names over the coming phases.

6. Compare with your current harness on your own tasks. Microsoft publishes no benchmark here.

What I would not conclude

• That the performance claims hold for your workload. The posts give no numbers.
• That this replaces your own agent loop. Microsoft's own table keeps hosted agents as the path when you need to change the loop.
• That it will stay free. The Foundry post says billing starts December 1.
• That the customer quotes prove anything general. They are two testimonials, one saying a solution ran "in less than a week".
• That switching harnesses is easy later. The post says switching after creation is not supported yet.

Frequently asked questions

What is the Microsoft Copilot harness?

Per Microsoft, a shared agent foundation, originally built by GitHub engineers, that powers Copilot products and is now offered through Foundry Agent Service in public preview.

What does Microsoft run and what do I run?

The Foundry post says Microsoft runs the agent loop and sandbox. You provide instructions, skills, tools and model choice.

How much does it cost?

Harness runtime is free before December 1, 2026. After that, the post lists sandbox rates of USD $0.0994 per vCPU-hour and $0.0118 per GiB-hour, plus separate charges for model inference and tools.

Can I switch an existing agent to it?

Not yet. The post says existing agents are unchanged and switching the harness after creation is not supported.

What is the difference from hosted agents?

Hosted agents let you bring your own framework, loop and runtime while Foundry manages hosting. The harness is the managed loop for agents you define declaratively.

Does Foundry recover a crashed long-running agent?

Per the post, it detects interruptions and restarts with the original input. Your code must save progress and avoid duplicate actions.

What happens to GitHub Copilot SDK?

It is becoming Microsoft Copilot SDK in phases. The changelog says no immediate changes are required.

Sources

All read on October 29, 2026.

• Introducing the Microsoft Copilot harness (Official Microsoft Blog, October 28, 2026)
• Introducing the Microsoft Copilot harness in Foundry Agent Service (Microsoft Foundry Blog, October 28, 2026)
• GitHub Copilot SDK is becoming Microsoft Copilot SDK (GitHub Changelog, October 28, 2026)

Not read: the Foundry documentation pages on supported regions, models, tools and preview limits, the quickstart, the samples, and the SDK repository.