Microsoft opens its Copilot agent harness as a managed Foundry service

Microsoft put the agent harness that runs GitHub Copilot into Foundry Agent Service. The service is in public preview. Microsoft said this in a developer blog post on Wednesday. The same harness now backs Copilot Studio, Copilot Code, the CLI and VS Code.

Microsoft runs the agent loop and the sandbox. The developer picks a supported model, writes a skill, connects tools and sends instructions. The idea is to let a team release a working agent without owning the runtime underneath it.

Microsoft says the harness manages four things developers would otherwise build. The first is reusable skills, written in Markdown. The second is tools from Foundry and Model Context Protocol, each with per-call authentication. The third is execution with files and subagents, where the main agent coordinates the result. The fourth is agent definitions in an azure.yaml file, kept in source control.

A quote in Microsoft's post is from Chris Evans, chief technology officer of Confluence Technologies. He says the team had a working agent in less than a week.

Pick a harness once

Two limits are in the post. The Copilot harness has to be picked when the agent is first created. The choice cannot be changed later, even in a new version. Microsoft also controls how much CPU and memory the agent gets, and how long it can sit idle.

A team that needs a different harness runs its own hosted agent. That path keeps Foundry hosting but drops the Copilot runtime. Bringing your own model is not supported inside the managed harness. The developer picks from supported models in Foundry.

The price starts in December

Microsoft said the harness runtime is free until 1 December 2026. Model inference still costs what it costs. The same holds for tools and connected services that bill separately. After 1 December Microsoft will charge $0.0994 per vCPU-hour for sandbox compute. Memory is billed at $0.0118 per GiB-hour. The post says prices may vary by region and agreement.

Microsoft also made long-running resilience generally available this week for hosted agents. The post describes it as recovery for work that takes hours or days. Foundry restarts the agent with its original input. The agent's own code decides what to save and how to resume. Conversation history on its own is not enough.

The trade is clear enough. A team that picks the Copilot harness gets Microsoft's shared runtime across every surface a GitHub user already touches. It loses the right to swap that runtime later. On a new project the constraint is cheap. On an agent that outlives its first design, the cost shows up when the harness no longer fits.

Source

Microsoft, Tina Schuchman and Vladimir Federov. Reporting: The New Stack, Amanda Caswell.