Copilot harness in Foundry: free until Dec 1, then metered

Microsoft wants to run your agent loop for you. On October 28 it put the Microsoft Copilot harness in Foundry Agent Service into public preview: you declare the agent, Microsoft operates the loop and the sandbox it works in. The runtime is free until December 1, 2026. After that, the meter starts — and the meter is the part worth reading.

What actually happened

The harness is a managed runtime for agents that reason, call tools and run long tasks. Microsoft's post lists four building blocks:

• Skills that package instructions, scripts and resources
• Foundry and MCP tools with configured authentication and approvals
• Execution and delegation, including file and code work and built-in subagents
• Source-controlled agent definitions, built in the Foundry portal or in code

In code, you set harness: type: microsoft_copilot_preview in an azure.yaml, attach a skills folder, deploy with azd deploy --all, and call it with azd ai agent invoke. You pick a supported Foundry model. Microsoft also made long-running resilience for hosted agents generally available: Foundry restarts interrupted work with its original input.

Preview limits matter: you can't switch an agent's harness after creation, and Foundry, not you, controls the sandbox's CPU, memory and idle settings.

Why it matters for your business

From December 1, sandbox compute bills at $0.0994 per vCPU-hour and $0.0118 per GiB-hour of memory, before model inference and any separately billed tools. Microsoft notes prices can vary by region and agreement. Those numbers look small. But agent cost is runtime times concurrency, and you don't control the idle settings. An agent that sits waiting on a slow API or a human approval is still a running sandbox.

So use the free month as a measurement window. Run a real workflow, log wall-clock time per task, and multiply it out against your monthly volume before December 1. Then you have a number to compare against a plain hosted agent or a script you own.

The other note is in Microsoft's own fine print: on restart, your agent must save checkpoints and prevent duplicate actions in external systems. "Resilient" means the task runs again. If that task sends an invoice, make sure it can't send two.

Sources: Microsoft Foundry Blog.