Microsoft Copilot Harness in Foundry: Setup and Pricing
Microsoft's new Foundry Copilot harness is in public preview. See setup, supported regions, production limits and December 1 runtime pricing.
Microsoft introduced the Copilot harness for Foundry Agent Service on October 28, 2026. It is a managed runtime for prompt agents that need more than a single model response: skills, tools, files, sandbox execution and bounded delegation. This is a public preview of a specific harness, not a blanket upgrade of existing Foundry agents or a generally available production feature.
What the managed harness actually changes
Foundry operates the agent loop and execution environment while the developer supplies a supported model, instructions, approved tools and procedures. The harness coordinates tool calls, maintains a conversation-scoped workspace, uses reusable skills and delegates bounded tasks to subagents. It also supports connected Foundry and MCP tools with their own authentication and approval requirements. This reduces orchestration code that a team must maintain, but it does not remove the need to design safe permissions, evaluate outputs or store durable business state.
How to test it
- Confirm your Foundry project uses a supported harness region and a deployed model with reasoning and tool-calling support.
- Create a new prompt agent in the current Foundry portal and select Use the Microsoft Copilot harness; existing agents cannot switch harnesses during the preview.
- Attach one small skill and the minimum tools required; review project roles, agent identity and tool connection permissions.
- Save, start a new conversation, run a bounded test and inspect the trace, tool activity and generated files.
- Measure task correctness, execution time and all model and tool charges before expanding the pilot.
Teams that keep definitions in source control can use the documented Azure Developer CLI route with harness type microsoft_copilot_preview. Supported SDKs can also create a prompt agent with this harness. Applications still invoke the agent through its Responses endpoint. Because the harness is selected at agent creation, a migration plan should assume a separate agent and a controlled comparison rather than an in-place switch.
Pricing changes on December 1
Published Copilot harness runtime rates
The pricing trap is treating the preview as a free AI agent. Microsoft says only harness runtime usage is free before December 1; inference and separately charged tools follow their own rates. After that date, compute and memory are metered separately. A useful pilot budget should record task success, model tokens, tool calls, execution duration and runtime consumption, not compare the harness compute price alone with a self-hosted application.
Preview limits that matter
- The harness preview has no service-level agreement and is not recommended for production workloads.
- Supported geography is narrower than general Foundry Agent Service; verify region, model and tool compatibility together.
- Bring-your-own-model is not supported for this harness preview.
- Foundry manages the sandbox CPU, memory and idle lifecycle. Conversation workspaces are not a durable output store.
- Built-in file, web, shell and subagent capabilities remain subject to service and organizational policies; shell access warrants careful review.
Managed harness versus hosted agents
Choose the right runtime ownership
| Use a prompt agent with Copilot harness when | Use a hosted agent when |
|---|---|
| You want managed multi-step execution, skills, files and supported tools. | You need your own framework, agent loop, middleware or dependencies. |
| The task fits the preview sandbox, model and region restrictions. | You need different runtime behavior or application-specific recovery logic. |
| You can test with bounded permissions in a preview-safe project. | You need a more customized execution architecture. |
For an internal research assistant, the harness could eliminate coordination work: define a reusable evidence procedure, connect approved sources and let the managed loop assemble a brief. For an agent that updates customers or financial systems, preview restrictions and the missing SLA change the risk calculation. In both designs, idempotent writes, audit logs and human checkpoints remain application responsibilities. Microsoft separately announced generally available long-running resilience for hosted agents; that is not a GA release of the Copilot harness.
Questions people ask
Is the Microsoft Copilot harness in Foundry production-ready?
No. Microsoft calls this harness a public preview without a service-level agreement and does not recommend preview features for production workloads. Foundry Agent Service itself is generally available, but that does not make this particular harness generally available. Test it in a non-production project, keep human approval for consequential actions, and check the current Microsoft Learn preview guidance before expanding use.
Can an existing Foundry prompt agent be switched to the Copilot harness?
No, not during the initial public preview. Microsoft says you must select the Copilot harness when creating a new prompt agent; publishing a new version of an existing agent does not change its harness. Create a separate test agent, select the harness in the Foundry portal or a supported SDK, and migrate instructions and tools only after verifying permissions and compatibility.
Is the Copilot harness free until December 1, 2026?
Only the harness runtime is free before December 1, 2026; model inference and separately billed tools or connected services can still cost money. Microsoft says sandbox compute will then cost $0.0994 per vCPU-hour and memory $0.0118 per GiB-hour, subject to region and agreement. These are runtime rates, not an all-inclusive price per task, so budget for model and tool usage separately.
Which Azure regions and models support the Copilot harness?
The initial preview supports Australia East, Canada Central, Central US, Germany West Central, Japan East, Sweden Central, UK South, West Europe and West US. Your model must support reasoning and tool calling and be available to prompt agents in that region. Bring-your-own-model is not supported. Check the current Microsoft Learn support table before creating a project because region and tool coverage may change.
The Copilot harness gives Foundry prompt agents more managed execution capabilities without making developers operate the entire agent loop. Public-preview status, creation-time setup, restricted regions and December 1 runtime billing are the main decisions today. Start with a low-risk measured pilot; choose hosted agents when custom framework or runtime control matters more than convenience.
Sources & useful resources
- Microsoft Foundry announcement, October 28, 2026— Launch, pricing, preview and hosted-agent comparison
- Microsoft Learn: Use Copilot harness with prompt agents— Setup, permissions, creation and preview limits
- Microsoft Learn: Foundry Agent Service regions and limits— Nine regions, supported model and BYOM constraints