Eigen RadarAI
Analysis

Microsoft opens Copilot’s execution layer to outside developers

Microsoft is making the infrastructure around its Copilot models available through Foundry and a downloadable software kit. The GitHub-built harness manages instructions, context, tools and execution while customers choose models and supply their own capabilities. Foundry offers managed operation for agents carrying out extended tasks. Microsoft says the shared foundation already underpins several of its Copilot products.

Artificial Intelligence··Night
In an imagined cream room, a red oval chain mechanism links a black cartridge, ceramic vessel and cord spool to a gearbox; a large crank reaches a slider on the lower track.

Copilot’s surrounding infrastructure becomes a service

Microsoft opened the execution infrastructure behind its Copilot assistants to developers through Foundry, its platform for building AI applications, on 28 October. Known as the Copilot harness, this layer organizes the instructions, context and tools around a model. GitHub, Microsoft’s software-development platform, built the foundation that Microsoft says it now uses across several of its own products.[1], [2]

An agent needs more than a model response when it must use tools and complete a sequence of steps. The harness provides the execution environment and controls that accompany those actions. Microsoft offers a managed cloud route through Foundry as well as a downloadable software development kit, allowing teams to build applications on the same underlying execution layer.[1]

Customers choose tools and models

Customers supply their instructions, reusable skills, tools and model choice while Microsoft operates the agent loop and sandbox. The loop manages repeated steps toward a task; the sandbox supplies an isolated place for execution. The company says the harness supports multiple model families and can route work according to the task, without requiring a different surrounding infrastructure for each model.[1]

The shared system supports the Model Context Protocol, a standard for connecting AI applications with tools and information, along with agent skills and plugins. Microsoft says other frameworks remain usable on its platforms. The announcement therefore gives teams another execution option while leaving them control over the capabilities that their application exposes to an agent and the models that perform its work.[1]

The shared foundation already runs Copilot products

Microsoft identifies Copilot Code, Copilot Studio, the GitHub Copilot app, its command-line interface and Visual Studio Code among the products using the harness. The company recently moved the runtime to the Rust programming language. It says updates to the common foundation can reach both its own products and customer-built agents, while managed operation removes part of the ongoing infrastructure maintenance from application teams.[1], [2]

References

  1. News sourceMicrosoftMicrosoft offers the Copilot agent harness through Foundry↩1↩2↩3↩4↩5
  2. News sourceUnite.AIMicrosoft opens the Copilot harness through Foundry↩1↩2