On this page
Microsoft’s Foundry developer blog dated October 28, 2026 (publish time corresponds to 00:45 Asia/Shanghai on October 29) introduces the Microsoft Copilot harness in Foundry Agent Service in public preview. That is a preview stage for the harness itself, not a generally available claim. Foundry Agent Service as a product is already generally available; the Copilot harness is a preview capability on top of it. The topic is Azure Foundry cloud-hosted agents, not the editor-side story in VS Code 1.143 Makes the Copilot Harness the Default for New Sessions.
Managed loop, agent definitions
Microsoft runs the agent loop and sandbox. Developers define agents declaratively in Foundry Agent Service and bring the knowledge, procedures, and tools for each agent. Microsoft’s company blog the same day says the same Copilot harness already underpins Copilot Code, Copilot Studio, the GitHub Copilot app, CLI, and VS Code, and is now offered to developers and enterprises through Microsoft Foundry.
Four preview building blocks
The public preview includes reusable skills; connected tools (supported Foundry tools and MCP, with explicit authentication and approvals); execution and delegation (files and code, plus built-in subagents); and source-controlled agent definitions (create and test in Foundry, or keep definitions next to application code). For a broader harness framing on this site: Same Model, Different Coding Harness: Pass Rate Barely Moves, the Bill Does. Another cloud-hosted agent runtime neighbor: AWS Opens Infinity, a Research Runtime for Extreme Agent Concurrency.
Same-day GA: hosted-agent resilience only
Separately, Microsoft announced general availability of long-running resilience for hosted agents. That GA covers recovery for long-running hosted agents; it is not a Copilot harness GA.
Billing and limits
Billing for the Copilot harness in Foundry Agent Service begins December 1, 2026, consumption-based. Harness runtime usage is free before December 1; model inference and separately billed tools and connected services still follow their own charges even during the preview. From December 1, sandbox compute and memory are listed at USD $0.0994 per vCPU-hour and $0.0118 per GiB-hour (prices may vary by region and agreement). Microsoft Learn lists the harness in a subset of prompt-agent regions and marks BYOM Not supported.
For the Copilot harness on Azure Foundry, DigIntoAI has not yet created an agent, run sandbox billing, or verified resilience recovery. Details follow the Foundry developer blog, Microsoft’s company blog, and Microsoft Learn. For reference only.