The system around the model
Microsoft is making its Copilot harness available through Foundry, according to the company’s announcement. The harness is the orchestration layer around a model: it supplies instructions, context, an execution environment and controls for an agent’s work.
Microsoft says the technology originated with GitHub’s engineering team and now powers several Copilot surfaces. Developers can download an SDK or use a cloud-managed version in Foundry. In the managed arrangement, Microsoft runs the agent loop and sandbox, while developers provide instructions, skills, tools and model choice.
Keeping model choice separate from orchestration
The announcement describes support for more than 20 frontier models and bring-your-own-key use. It also names MCP, agent skills and plugins as supported standards or extension mechanisms. The intended benefit is that changing a model does not require replacing the surrounding agent system.
That distinction matters for teams evaluating a long-running workflow. Model quality is only one variable; tool access, context handling and execution controls also affect the result. A common harness can make those surrounding choices more consistent when comparing models, although it cannot make every model equally capable.
Evaluate a complete workflow
Microsoft reports performance improvements following a move of the runtime to Rust, but the announcement does not provide a universal latency or cost figure for customer applications. Those claims should remain attributed to the company rather than become an assumed saving in a project budget.
A practical evaluation would use one representative task with explicit completion criteria and approved tools. Measure completed work, failures and total cost under the intended model configuration. The product offers a managed starting point and continued harness updates; teams still need to decide which actions their own agent may take and how they will review its results.