Microsoft has announced a public preview of the Copilot harness in Foundry Agent Service, according to the Azure Blog. It is a managed execution layer for agents defined declaratively in Foundry. Microsoft runs the agent loop and the sandbox. The customer supplies the knowledge, procedures and tools.
What the preview includes
The preview combines four building blocks:
- Reusable skills: instructions, scripts and supporting files packaged as a procedure an agent can apply across tasks.
- Connected tools: supported Foundry tools and Model Context Protocol (MCP) tools, with authentication and approvals configured explicitly.
- Execution and delegation: agents can work with files and code and hand bounded tasks to built-in subagents, while the main agent assembles the result.
- Source-controlled definitions: agents can be created and tested in Foundry, or kept in the same repository as the application code.
The harness tracks working context as a task develops and coordinates model and tool steps within the limits and policies you set. You pick a supported Foundry model without rewriting the orchestration around it.
Microsoft's example is a research-brief agent. A skill file describes the method, and a YAML definition attaches web search and a code interpreter. Deployment and invocation go through the Azure Developer CLI. The Foundry portal, a Visual Studio Code toolkit and a GitHub Copilot app plugin are other ways to start.
Long-running hosted agents
Separately, long-running resilience for hosted agents is now generally available. Hosted agents let a team bring its own framework and loop while Foundry handles hosting. When an interruption hits work that takes hours or days, Foundry detects it and restarts the job with its original input.
Resuming is the developer's job. The code must save progress and intermediate results, because conversation history alone cannot restore work in progress. If the agent writes to another system, the workflow must also check for existing output so a restart does not repeat an action.
Choosing a level
Microsoft describes three routes. Prompt agents without the harness suit tasks where instructions and supported tools are enough. Prompt agents with the harness suit work that needs skills, files, subagents and heavier tool coordination. Hosted agents suit teams that must control the framework, loop, middleware or runtime.
Limits and cost
During the preview, the harness is chosen at the moment the agent is made. Agents that already exist remain as they are, and changing the harness once an agent has been created is not yet available. Foundry configures the sandbox's CPU, memory, and idle behavior.
Billing for the harness starts December 1, 2026, on a consumption basis. Runtime use is free before that date, but model inference and separately billed tools and connected services are charged as usual. From December 1, sandbox compute costs USD $0.0994 per vCPU-hour and memory costs $0.0118 per GiB-hour. Prices may vary by region and agreement.
Two customers are quoted. Confluence Technologies' CTO says a solution was running in under a week. Fashable's CEO says its agents can discover, engage and calculate while commercially binding decisions stay under human control.