Microsoft opens Copilot Harness to enterprise developers

Moving beyond the model

Microsoft has announced that it is opening the Copilot Harness to external enterprise developers during the GitHub Universe 2026 event. This software acts as the orchestration engine that previously powered GitHub Copilot internally. It is now available in public preview through the Foundry Agent Service.

The Copilot Harness is not an AI model itself but rather the infrastructure layer that sits above the models. It manages what an AI agent knows, which tools it can access, how long it can run, and what actions it is permitted to take. It handles complex tasks like instruction injection, context window management, and session sandboxing.

By providing this layer to the public, Microsoft is allowing teams to build their own agents on the same infrastructure used for Copilot's most complex coding tasks. This move marks the first time external developers can build directly on the harness. Previously, the technology was only available inside Microsoft's own products.

Model-agnostic flexibility

A key feature of the Copilot Harness is its design to work independently of any specific AI provider. It accommodates more than 20 leading models from the OpenAI and Anthropic families, including both the Claude and GPT series. Developers can also use bring-your-own-model configurations to suit specific needs.

The system includes a multi-model routing layer that can select the best model dynamically based on the complexity of a task or specific latency requirements. Organizations can also set fixed routing preferences at the policy level. This allows companies to switch models without having to rewrite their entire agent code as new models emerge.

Because the harness manages state, memory, and policy at the orchestration level, it simplifies governance for large organizations. Instead of configuring rules separately for every different model family, a team only needs to configure the harness once. This ensures consistent behavior across a diverse fleet of AI tools.

Performance and connectivity

To improve performance for enterprise use cases, Microsoft ported the harness runtime to Rust. This engineering choice reduces the overhead required to initialize sessions and improves throughput for long-running tasks. This is particularly useful for workflows like document analysis or cross-system data retrieval that require many steps.

The platform also includes long-running resilience capabilities to handle interruptions. The Foundry Agent Service can detect when a session is interrupted and restart the work without losing progress. This transition allows the harness to support workflows that may last for hours or even days rather than just short interactions.

Additionally, the harness natively supports the Model Context Protocol (MCP), an open standard for connecting agents to external data. Because it follows this standard, agents can connect to over 10,000 public MCP servers. This includes integrations with databases, calendars, and file systems without requiring developers to write bespoke connectors for every new tool.

The orchestration trade-off

Industry analysts suggest that the primary advantage in enterprise AI is moving away from the quality of individual models toward the infrastructure that coordinates them. Microsoft's move is seen as a way to convert vendor concentration risk into a platform monetization opportunity. It allows for model diversity at the application layer while maintaining developer dependency at the runtime layer.

For enterprise teams, choosing the Copilot Harness involves a specific architectural tradeoff. On one hand, it provides production-grade sandboxing and proven infrastructure at scale. On the other hand, it means accepting Microsoft's specific session policies, pricing models, and routing decisions as the foundation of the agent environment.

Organizations that require full control over how routing decisions are made or how policy is enforced may find themselves working around the platform's constraints. However, for many standard enterprise use cases, the benefit of using a production-ready system outweighs the limitations of the configuration. Teams must evaluate these dependencies carefully, especially in regulated industries with strict audit requirements.

Pricing and access

The Copilot Harness is currently available in public preview at no charge through November 30, 2026. This period allows developers to test and run agents without incurring costs. Access is provided through the Foundry portal, the VS Code Foundry Toolkit extension, the azd command-line interface, and the GitHub Copilot app.

Starting December 1, the platform will move to a consumption-based billing model. Under this structure, sandbox compute runs at $0.0994 per vCPU-hour, and memory costs $0.0118 per GiB-hour. There are no per-seat or per-model-call charges at the harness level, meaning costs only track the actual compute and memory used by active sessions.

Developers interested in the platform can attend GitHub Universe 2026 for additional sessions on the harness and the broader Foundry platform. The event continues through October 29 at the Fort Mason Center in San Francisco. These sessions will provide deeper technical details on how to integrate the harness into existing enterprise workflows.