Microsoft will run the agent loop for you, with billing from Dec. 1

Microsoft has put the Copilot harness into public preview in Foundry Agent Service. According to the official announcement, Microsoft runs the agent loop and sandbox, and developers supply skills, tools and knowledge. Runtime usage is free until December 1, 2026.

Skills, tools and subagents in a managed sandbox

The harness is meant for agents defined declaratively in Foundry. Microsoft describes four building blocks: reusable skills (instructions, scripts and resources packaged as procedures), connected tools, execution with delegation, and agent definitions kept under source control.

Tools can come from Foundry or from Model Context Protocol (MCP) servers, with authentication and approvals configured explicitly. Agents can work with files and code, and can hand bounded tasks to built-in subagents while a main agent assembles the result. The developer picks a supported Foundry model, and Microsoft says the orchestration does not need to be rewritten around it.

The example in the post uses the preview Azure Developer CLI (azd) surface. A skill lives in skills/research-brief/SKILL.md. The agent is described in azure.yaml with the harness type microsoft_copilot_preview, then shipped with azd deploy --all and called with azd ai agent invoke. Agents can also be created in the Foundry portal, in Foundry Toolkit for Visual Studio Code, or through the Microsoft Foundry plugin in the GitHub Copilot app.

Microsoft also says long-running resilience for hosted agents becomes generally available this week. Foundry detects an interruption and restarts the work with its original input. Saving progress is the developer’s job, and so is avoiding duplicate actions, such as creating the same report twice in another system.

Free runtime for now, then hourly sandbox billing

Billing begins on December 1, 2026, on a consumption basis. From then, sandbox compute costs $0.0994 per vCPU-hour and memory costs $0.0118 per GiB-hour, and prices may vary by region and agreement. Model inference and separately billed tools and connected services are charged regardless.

As a rough illustration (our arithmetic, not Microsoft’s), a sandbox with 4 vCPUs and 16 GiB would come to about $0.59 per hour at those rates. Microsoft says Foundry controls CPU, memory and idle settings during the preview, so developers cannot pick the size themselves yet.

The preview has other limits. The harness must be selected when an agent is created, existing agents are unchanged, and switching the harness later is not supported yet. The post points to a separate page for supported regions, models, tools and preview limits, which it does not summarize.

Two customers are quoted. Confluence Technologies CTO Chris Evans said his team had a solution running in less than a week. Fashable CEO Orlando Ribas Fernandes said commercially binding sourcing decisions should stay under human control. Both are Microsoft-selected statements and are not independent evaluations.

Where it sits next to hosted agents and rivals

Microsoft’s own table shows three levels. Plain prompt agents suit tasks where instructions and supported tools are enough. Prompt agents with the harness add skills, files, subagents and more complex tool coordination. Hosted agents suit teams that need their own framework, loop or middleware while Foundry handles hosting.

That third tier already has a history. Microsoft launched hosted agents in Foundry Agent Service in public preview in April 2026, with sandboxed sessions, according to a Foundry developer blog post and press coverage of it. The harness therefore adds a ready-made loop on top of a runtime Microsoft had already shipped.

It also fits a stated multi-framework position. In a June 2 Microsoft blog post on its agent platform, the company said Foundry accepts agents built on the GitHub Copilot SDK, the Claude Agent SDK, LangGraph and custom harnesses, and that Foundry can tune harnesses alongside models and tools. Offering an in-house harness while hosting rival ones lets Microsoft compete at the loop layer without closing the platform.

The same Copilot branding runs across the company’s agent products. Our earlier coverage of the GitHub Copilot app’s Projects and computer use mode noted that plan inclusion and pricing were not specified there either. Microsoft’s post refers to a wider “unified harness strategy”, but the page itself does not detail how the Foundry harness relates to the one inside the GitHub Copilot app.

Much is still unestablished. The post gives no benchmarks, no latency or reliability numbers and no list of supported models in its text. The customer time-to-build claims are unverified. Developers who want to try the harness can start in the Foundry portal or with azd, and have until December 1 to test it with no runtime charge.