Microsoft on Wednesday opened the orchestration engine that has powered GitHub Copilot to external enterprise developers, announcing at GitHub Universe 2026 that the Microsoft Copilot Harness is now available in public preview through the Foundry Agent Service. The move gives any enterprise development team access to the same sandboxed compute, policy controls, and multi-model routing infrastructure that Microsoft has used internally to run Copilot’s most complex multi-step coding tasks — at no charge through November 30, 2026, with consumption-based billing starting December 1.

The announcement, made jointly by Corporate Vice President Tina Schuchman of Foundry and GitHub CTO Vladimir Fedorov at the Fort Mason Center in San Francisco, marks the first time external developers can build on the harness directly. Until now, the harness was available only inside Microsoft’s own products; it reached general availability in Copilot Studio in August 2026.

What the Copilot Harness Actually Does

The Copilot Harness is not a model. It is the layer above the model — the infrastructure that manages what an AI agent knows, what tools it can reach, how long it can work, and what it is allowed to do. In technical terms, it functions as an orchestration wrapper: it handles instruction injection, context window management, tool call execution, session sandboxing, and enterprise policy enforcement, then routes the work to whichever underlying AI model is best suited for the task.

The distinction matters because it means the harness is deliberately model-agnostic. It supports more than 20 frontier models from the OpenAI and Anthropic model families, including models from both the Claude and GPT lineages, as well as bring-your-own-model configurations. Enterprises can set routing preferences at the policy level or allow the harness’s built-in multi-model routing to select dynamically based on task complexity and latency requirements.

Since the harness manages state, memory, and policy at the orchestration level rather than delegating those responsibilities to individual models, organizations applying governance rules to AI agents need only configure the harness once — not separately for each model family they use.

Rust Runtime Brings Performance Gains

Microsoft ported the harness runtime to Rust ahead of the external release, a choice that reduced session initialization overhead and improved throughput for long-running agent tasks. That engineering work matters practically: enterprise agents running multi-step workflows — document analysis, code review pipelines, cross-system data retrieval — can spend significant time on runtime overhead between task steps, and a lower-latency session layer reduces that cost at scale.

The platform also ships with long-running resilience capabilities, enabling the Foundry Agent Service to detect session interruptions and restart work without losing progress. This moves the harness from a tool suited for short interactive exchanges to one capable of sustaining the hour-long or day-long workflows that enterprise automation increasingly demands.

Model Context Protocol Integration Means Immediate Ecosystem Access

The harness natively supports the Model Context Protocol, the open standard that Anthropic introduced in November 2024 and subsequently donated to the Linux Foundation’s Agentic AI Foundation for open governance. By the end of 2025, more than 10,000 public MCP servers had been published, covering integrations with databases, file systems, calendars, development tools, and enterprise APIs.

Because MCP is a client-server standard rather than a vendor-specific protocol, agents built on the Copilot Harness can connect to any MCP-compatible tool without custom integration work. A developer building a code review agent, for example, can attach it to a GitHub MCP server, a Jira MCP server, and an internal ticketing system’s MCP server using the same connection pattern, rather than writing bespoke connectors for each. The harness ships alongside a library of reusable agent skills — packaged instruction sets and tool configurations — and a plugin architecture that allows skills to be shared and versioned across an organization’s agents.

What the Orchestration-Layer Strategy Means for Enterprise Teams

The Copilot Harness launch is being read by enterprise AI architects as confirmation of a pattern that has been developing across 2026: the competitive moat in enterprise AI is shifting from model quality to orchestration infrastructure. As one industry analysis put it, Microsoft’s move to an orchestration layer “converts vendor concentration risk into a platform monetization opportunity” — enabling model diversity at the application layer while capturing developer dependency at the runtime layer.

That framing surfaces a genuine architectural tradeoff for enterprise teams evaluating the harness. On one side: production-grade sandboxing, MCP interoperability, multi-model flexibility, and the credibility of infrastructure already proven at GitHub Copilot’s scale. On the other: building on the harness means accepting Microsoft’s session policies, pricing model, and routing decisions as foundational constraints on how agents behave. Organizations that want to switch routing logic, change policy enforcement mechanisms, or run the harness in a fully self-hosted configuration will find themselves working around the platform rather than with it.

That is not a reason to avoid the harness — it is a reason to understand what the dependency means before committing to it as infrastructure, particularly for use cases where agent behavior is subject to regulatory audit or contractual compliance requirements.

How Much Does It Cost and How Can Developers Get Started?

The Copilot Harness is available in public preview at no charge through November 30, 2026. Starting December 1, consumption-based billing activates: sandbox compute runs at $0.0994 per vCPU-hour, and memory at $0.0118 per GiB-hour. There are no per-seat or per-model-call charges at the harness level; costs track only the compute and memory consumed by active agent sessions.

Access is available through the Foundry portal, VS Code’s Foundry Toolkit extension, the azd command-line interface, or the GitHub Copilot app. Developers and organizations interested in building on the harness before billing begins have until the end of November to run agents in preview without incurring charges.

GitHub Universe 2026 continues through Thursday, October 29, with additional sessions covering the harness and the broader Foundry platform at the Fort Mason Center in San Francisco.

Frequently Asked Questions

What is the Microsoft Copilot Harness, and how is it different from Copilot itself?

The Microsoft Copilot Harness is the orchestration layer that sits between enterprise applications and AI models. While Copilot is the end-user-facing product — the assistant embedded in GitHub, Microsoft 365, and other tools — the harness is the infrastructure underneath it that manages instructions, context, tool access, session sandboxing, and policy enforcement. What Microsoft announced at GitHub Universe 2026 is that external developers can now build their own agents on this infrastructure through the Foundry Agent Service, rather than being limited to agents Microsoft itself has built.

What AI models does the Copilot Harness support?

The harness supports more than 20 frontier models, including models from OpenAI’s GPT family and Anthropic’s Claude family, as well as bring-your-own-model configurations. Its multi-model routing layer selects dynamically between models based on task characteristics, or enterprises can configure fixed routing policies for compliance or cost-optimization purposes. The harness is designed to be model-agnostic: as new frontier models become available, they can be added to the supported pool without requiring developers to rewrite their agent code.

What is the tradeoff of building on the Copilot Harness rather than a self-managed orchestration stack?

The harness delivers production-grade infrastructure — proven at GitHub Copilot’s scale, with MCP interoperability out of the box, Rust-powered session performance, and built-in long-running resilience — faster than most engineering teams can build equivalent infrastructure themselves. The tradeoff is dependency: organizations building on the harness accept Microsoft’s session policies, routing architecture, and pricing structure as the foundation of their agent environment. Teams that require full control over how routing decisions are made, how session state is persisted, or how policy is enforced will need to build around those constraints. For most enterprise use cases, the production-readiness gain outweighs the configuration limits; for regulated industries with strict agent auditability requirements, the constraint analysis warrants more scrutiny before committing to the platform.

How does Model Context Protocol (MCP) support expand what these agents can do?

MCP is an open standard, originally introduced by Anthropic in November 2024 and now governed by the Linux Foundation’s Agentic AI Foundation, that defines a consistent way for AI agents to connect to external tools and data sources. Because the Copilot Harness natively supports MCP, agents built on it can immediately connect to the 10,000-plus MCP servers without custom integration work. In practice, this means a developer building an enterprise agent can attach it to calendar tools, code repositories, internal databases, and third-party APIs through a single, standard connection pattern — rather than writing and maintaining separate connectors for each data source.