Microsoft Hands Developers Its Copilot Agent Foundation: Copilot Harness Opens Through Foundry

What happened

What opens up is the layer around the model: Microsoft defines the harness as the orchestration layer around a model, giving an agent its instructions, context, execution environment and controls, with support for the Model Context Protocol, agent skills and agent plugins; the company says building and maintaining a cross-model orchestration layer in-house is costly, and with this release developers can use the same foundation as Microsoft and GitHub's own products.

Multi-model support and performance are the two points in the announcement: the harness supports more than 20 frontier models plus bring-your-own-key setups, routing work to the right model or combination of models for each task; the runtime was recently ported to Rust, and Microsoft says session overhead and time to the first token fell, without giving specific figures in the announcement; the same harness has been standardized inside Microsoft across all Copilot experiences.

The division of labor and billing are set on the Foundry side: developers can download the software development kit or use the cloud-managed version through Foundry, with Microsoft running the agent loop and sandbox, governance and optimization built in, while customers supply instructions, skills, tools and model choice, and the platform still accepts other harnesses and frameworks; the public preview in Foundry Agent Service requires selecting the harness when creating an agent, switching after creation is not supported yet, and Foundry manages the execution environment's CPU, memory and idle settings; harness runtime is free before Dec 1, after which sandbox compute is billed at $0.0994 per vCPU-hour and memory at $0.0118 per GiB-hour, with model inference billed separately.

Key facts

1. Announced by: Microsoft announced on its official blog on Oct 28 that the Copilot harness built by GitHub engineers is available through Microsoft Foundry.
2. Availability: public preview in Foundry Agent Service, with long-running resilience for hosted agents generally available at the same time.
3. Existing use: the harness already powers Copilot Code, Copilot Studio, the GitHub Copilot app, the command line and VS Code, and has been standardized inside Microsoft.
4. Definition: the harness is the layer around a model, handling instructions, context, execution environment and controls, with support for the Model Context Protocol, agent skills and agent plugins.
5. Model support: more than 20 frontier models plus bring-your-own-key setups, routing work to the right model or combination of models for each task.
6. Performance: the runtime was recently ported to Rust, and Microsoft says session overhead and time to the first token fell, with no specific figures in the announcement.
7. Division of labor: Microsoft runs the agent loop and sandbox, with governance and optimization built in; customers supply instructions, skills, tools and model choice; the platform still accepts other harnesses and frameworks.
8. Preview limits: the harness must be selected when creating an agent, switching after creation is not supported yet; Foundry manages the execution environment's CPU, memory and idle settings.
9. Billing: harness runtime is free before Dec 1, after which sandbox compute is billed at $0.0994 per vCPU-hour and memory at $0.0118 per GiB-hour, with model inference billed separately.