Microsoft’s Copilot harness shifts agent plumbing into Foundry

Microsoft announced its Copilot harness on October 28, extending the orchestration behind its agent products into Microsoft Foundry. A harness is the system around an AI model: it supplies working context, tools, execution and controls. The distinction matters for organizations evaluating agents because choosing a model alone does not provide the machinery needed to carry out a multistep task. Microsoft is offering that surrounding system as a managed option. ^[1]

Foundry’s implementation is a public preview. Its documentation says Microsoft runs the agent loop and sandbox while customers provide procedures, knowledge and tools. The initial components include reusable skills, authenticated tool connections, bounded subagents and source-controlled definitions. Developers must select the harness when creating an agent; switching it afterward is not yet supported. That makes the execution approach an early design choice, rather than a setting to assume can be changed later. ^[2]

The pricing distinction is equally concrete. Microsoft says harness runtime use is free before December 1, but model inference, separately billed tools and connected services retain their own charges. Consumption billing for the runtime begins December 1. Teams should therefore price the whole workflow rather than equate a temporary runtime allowance with a free agent. The announcement does not demonstrate the cost or reliability of a particular customer’s application. ^[2]

A downloadable SDK provides another route. The inspected repository still uses GitHub Copilot naming while describing a phased move to the Microsoft organization. It says existing package names and namespaces may remain during the transition and requires no immediate changes from existing users. That is useful migration context: a new umbrella name does not mean applications should immediately replace working imports. ^[3]

For standard SDK use, the repository says a Copilot subscription is required; bringing a supported provider’s own API key is an alternative. Prompts count toward the relevant usage allowance. The SDK manages a CLI server process through JSON-RPC, which is a different deployment arrangement from a Foundry-managed service. Developers should decide which environment and billing route they intend to operate before treating the two entry points as interchangeable. ^[3]

About this item: News report. No first-hand reporting or independent testing is claimed. Corrections policy.