Microsoft announced the expansion of access to the GitHub Copilot architecture via the Microsoft Foundry platform, through the launch of the Foundry Agent Service in public preview. This step aims to grant external developers access to the same orchestration infrastructure used by the company's Copilot products to build and deploy autonomous AI agents.
What does the Copilot harness service offer?
The new service provides a fully managed environment that handles model and tool invocations, tracks context, and ensures agents adhere to defined policies. Key features include:
- Sandboxing: Secure execution of agents within a Microsoft-managed environment.
- Reusable skills: The ability to package instructions, scripts, and resources to support agents.
- MCP protocol integration: Support for integrating external tools with authentication and consent requirements.
- Sub-agent support: The ability of the main agent to delegate tasks to other agents and coordinate results.
Developers can define agents using azure.yaml files and specify models, tools, and instructions, then deploy them via the Azure Developer CLI (azd).
What changes practically for developers?
Despite the ease of getting started, Microsoft imposes technical restrictions that may affect the flexibility of long-term projects:
- Lock-in: Once the Copilot harness is chosen when creating the agent, this choice cannot be changed later, even when new updates are released.
- Limited environment control: Microsoft controls processing resources (CPU) and memory, in addition to allowed idle periods.
- Model restrictions: Developers are limited to choosing compatible models available via Foundry, and bringing your own model (BYOM) is currently not allowed within this specific architecture.
For teams requiring greater control, Microsoft offers the Foundry hosted agents option, which allows the use of custom frameworks while benefiting from hosting services, including flexibility for long-running operations that may take hours or days.
Availability and pricing
The service is currently available for free during the preview period until December 1, 2026, with model inference and connected service costs continuing separately. After the free period ends, Microsoft will adopt a consumption-based pricing model (vCPU-hour and GiB-hour), with prices varying based on geographic region and customer agreements.