Microsoft opens Copilot harness to Foundry for multi-model agents

Expanding the Copilot Infrastructure

Microsoft announced on October 28, 2026, that it is opening its Copilot harness to the Microsoft Foundry platform. This move provides a managed runtime for developers who want to build multi-model agents and integrate complex tool use into their applications. The harness is the same underlying technology that currently powers major products like Copilot Code, Copilot Studio, the GitHub Copilot app, the CLI, and VS Code.

By making this infrastructure available in Foundry, Microsoft aims to help enterprise customers build agents capable of reasoning and completing long-running tasks. The company wants to ensure that developers can build these systems without having to switch their entire technology stack or maintain a custom router. This availability allows for a more unified approach to deploying AI agents that require high levels of coordination.

Microsoft describes this move as an expansion of its platform rather than a replacement for existing tools. Customers who use Foundry will have the option to run the harness as a fully managed service in the cloud. Alternatively, they can download the SDK if they prefer to manage the execution themselves while continuing to use other frameworks if they choose.

How the harness functions

The Copilot harness acts as the orchestration layer that sits around an AI model. It is responsible for supplying the necessary instructions, providing the relevant context, and managing the execution environment. It also handles the controls required to keep the AI's actions within specific boundaries. Because the field of AI agents is moving so quickly, Microsoft notes that maintaining a custom router can be difficult for many companies.

A primary challenge for developers is managing the continuous balance of performance, quality, and cost. The Copilot harness is designed to handle these complexities automatically by routing work to the correct model or group of models for a specific task. It was originally built by GitHub engineers and has been standardized across various Copilot experiences to ensure consistency in how the AI interacts with users.

To stay compatible with the broader ecosystem, the harness supports open standards such as agent skills and agent plugins. It also supports MCP, which is a standard for connecting models to data and tools. This ensures that the harness can interact with a wide variety of external systems and specialized capabilities without requiring custom code for every single integration.

Technical upgrades and performance

Microsoft recently updated the harness runtime by porting it to Rust, a systems programming language known for its speed and safety. The company claims this transition improved overall performance and reduced both session overhead and the time-to-first token. While Microsoft did not provide independent benchmarks to prove these claims, the update is intended to make the system more efficient for high-demand environments.

The system is designed to be highly versatile, supporting more than 20 frontier models. It also includes bring-your-own-key support, which allows organizations to use their own API keys for the models they choose to run. This flexibility is important for companies that have specific requirements regarding which models they use and how they manage their data and costs.

The harness is described as being battle-tested across millions of users and hundreds of real-world use cases. By using a system that has already been refined in production environments, developers can avoid many of the common pitfalls of building agentic workflows from scratch. The goal is to provide a stronger system around the model that preserves the user's choice of which specific AI to use.

Strategic considerations for users

While the new Foundry availability is a significant expansion, Microsoft offers specific guidance on when to use the harness. The company notes that the harness is best suited for longer, context-heavy work that requires complex reasoning. For simpler tasks, using the harness might actually add unnecessary latency and increase the consumption of Copilot Credits.

Because of this, Microsoft recommends that developers evaluate their specific scenarios individually. They advise against automatically migrating every single workflow to the new harness without testing how it affects the user experience. The goal is to ensure that the complexity of the orchestration layer provides a tangible benefit to the specific task being performed by the agent.

Privacy and security remain central to this rollout for enterprise customers. Microsoft clarified that Copilot Studio agent data is not sent to the GitHub Copilot service. Furthermore, all existing commitments regarding privacy, security, compliance, and data residency continue to apply to the Foundry customers using these tools. This ensures that large organizations can adopt the new capabilities without compromising their data protection standards.

What comes next for Foundry

The release of the harness in Foundry follows a series of updates from late 2026 regarding connected agents and Copilot surface updates. These announcements suggest a broader trend toward making Microsoft's internal AI orchestration tools more accessible to the public. By providing both a managed cloud service and an SDK, Microsoft is positioning Foundry as a hub for complex agent development.

Developers can now begin exploring the SDK to see how the harness handles multi-model routing and tool use. The availability of these tools means that the barrier to entry for building