# Microsoft 通过 Foundry 开放 Copilot harness,支持多模型 Agent 开发
> 原标题:Introducing the Microsoft Copilot harness: Unlocking the potential of truly multi-model agentic work
- 来源:Microsoft:Official Blog
- 发布时间:2026-10-28T16:51:35.000Z
- AIHOT:https://aihot.news/items/tdp7qmrh74fdfarxbi4jaoc6v
- 原文:https://blogs.microsoft.com/blog/2026/10/28/introducing-the-microsoft-copilot-harness-unlocking-the-potential-of-truly-multi-model-agentic-work/
## 摘要
Microsoft 宣布将 GitHub 团队打造的 Copilot harness 通过 Microsoft Foundry 向开发者和企业开放,该 harness 已驱动 Copilot Code、Copilot Studio、GitHub Copilot app 和 CLI 等产品。
## 推荐理由
微软官方说明 Copilot harness 的能力和开放方式,开发者可据此评估是否用它托管多模型 Agent 运行时。
## 正文 · 中文译文
如果你想参与竞争,就需要支持多模型。而如果你想充分利用不同 AI 模型所提供的一切,就需要一个出色的 harness。但构建一个能在多个模型和多个界面之间无缝运行的 harness 极其困难。你需要解决性能、质量和成本问题。然后你还得维护它,这更难。
幸运的是,GitHub 工程团队构建了[一个业界一流、真正的多模型 harness](https://benchmarks.github.com/),我们已在全公司范围内采用。Microsoft Copilot harness 为 Copilot Code、Copilot Studio、GitHub Copilot 应用、CLI、VS Code 等提供支持。今天,我们通过 [Microsoft Foundry](https://devblogs.microsoft.com/foundry/microsoft-copilot-harness-in-foundry) 提供这一 harness。这为开发者和企业客户提供了一种完全托管的方式,用来构建能够推理、使用工具并完成复杂、长时间运行任务的智能体。你可以为任务选用合适的模型,而无需切换 harness 或技术栈的其余部分。这释放了你的时间、精力和资源,让你专注于自己的核心目标。
### **什么是 harness,为什么它很难?**
如果模型提供智能,那么 harness 就是模型周围的一切。它是编排层,为智能体提供正确的指令、上下文、执行环境和控制,以完成有意义的工作。一个名副其实的 harness 支持 MCP 和 agent skills 等开放标准,以及 agent 插件。
这是一个快速发展的领域。今天的前沿模型明天可能就会落后。而构建自己的 harness 或模型路由器来跟上进度,可能既痛苦又缓慢。现在,你可以基于 Microsoft 和 GitHub 在我们自己的产品中交付的同一个 harness 进行构建,而且无需维护它。
我们投入科学和工程工作,持续改进和演进 harness,确保在一系列模型和工作负载上实现高质量、高效和低成本的性能。这些收益会为我们以及使用该 harness 构建和运行自己智能体的客户带来复利效应。改进 harness,它触及的每个界面都会受益。
### **诞生于 GitHub,经过企业级规模实战检验**
Copilot harness 由打造我们[多模型编排](https://github.blog/ai-and-ml/github-copilot/project-hydrafusion-frontier-quality-via-multi-model-orchestration/)等新技术的同一批 GitHub 工程师构建。而这些技术之所以成为可能,正是因为我们构建了自己的 harness。它被设计为快速、token 高效且可预测,是我们最初的高质量 harness。它在要求最高的受众——开发者——中获得了最高的满意度。这就是为什么我们要将所有 Copilot 体验标准化在尽可能最强大的基础上。
该 harness 旨在[跨模型系列运行](https://x.com/ArtificialAnlys/status/2115477376981508419),支持 20 多个前沿模型以及自带密钥,因此你可以为合适的任务选用合适的模型。无需自行跟进和评估新模型,依靠我们根据任务智能地将工作路由到合适的一个或多个模型。
我们的 harness 灵活且可配置,适用于任何类型的应用、智能体或服务。我们最近[将 harness 运行时移植到 Rust](https://github.blog/ai-and-ml/generative-ai/migrating-the-github-copilot-runtime-to-rust-using-copilot/),从而实现了业界领先的性能,并减少了会话开销和首 token 时间。而且,由于 Copilot harness 为 Microsoft 各界面提供支持,它已经过实战检验,并在数百万用户和数百个真实用例的规模上证明了其性能。
统一的 harness 让我们能够更快地交付、更快地学习,并持续改进 Copilot。而且,由于你在 Foundry 中构建的智能体可以运行在我们于 Microsoft 全面采用的同一 harness 上,我们对 harness 的持续改进也将惠及你的智能体。
### **亲自试一试**
[SDK 已可供下载](https://github.com/github/copilot-sdk),你也可以在云端运行该 harness,全部由 Foundry 托管。我们的平台保持开放与互操作,因此你依然可以引入任何你想要的 harness 和框架。
借助 Copilot harness,Microsoft 负责运行智能体循环和沙箱,并从第一天起内置治理与优化。你只需提供指令、技能、工具和模型选择。
这是朝着我们愿景迈出的重要一步:围绕模型构建更强大的系统,同时保留选择权。它代表了 GitHub 的精华,并将其带到 Microsoft 平台及更广阔的领域。
我们将持续推动 harness 向前发展,让你能够专注于只有你才能构建的东西。
文章 [Introducing the Microsoft Copilot harness: Unlocking the potential of truly multi-model agentic work](https://blogs.microsoft.com/blog/2026/10/28/introducing-the-microsoft-copilot-harness-unlocking-the-potential-of-truly-multi-model-agentic-work/) 首次发布于 [The Official Microsoft Blog](https://blogs.microsoft.com/)。
## 正文 · 原文
If you want to compete, you need to be multi-model. And if you want to take advantage of everything that different AI models have to offer, you need a great harness. But building one that works seamlessly across multiple models and surfaces is incredibly hard. You need to solve for performance, quality, and cost. And then you have to maintain it, which is even harder.
Luckily, the GitHub engineering team built [a best-in-class, truly multi-model harness](https://benchmarks.github.com/), which we’ve adopted company-wide. The Microsoft Copilot harness powers Copilot Code, Copilot Studio, the GitHub Copilot app, the CLI, VS Code, and more. And today, we’re making that harness available through [Microsoft Foundry](https://devblogs.microsoft.com/foundry/microsoft-copilot-harness-in-foundry). That gives developers and enterprise customers a fully managed way to build agents that reason, use tools, and complete complex, long-running tasks. You get the right model for the job without switching the harness or the rest of your tech stack. That frees up your time, energy, and resources to focus on your bottom line.
### **What’s a harness, and why is it hard?**
If the model supplies the intelligence, the harness is everything around the model. It’s the orchestration layer that gives an agent the right instructions, context, execution environment, and controls to accomplish meaningful work. And a harness worth its salt has support for open standards like MCP and agent skills, as well as agent plugins.
This is a fast-moving space. A frontier model today might fall behind tomorrow. And building your own harness or model router to keep up can be painful and slow. Now, you can build on the same harness Microsoft and GitHub ship in our own products, and you don’t have to maintain it.
We invest in science and engineering work to continually improve and evolve the harness and ensure high-quality, efficient, and low-cost performance across a range of models and workloads. The benefits compound for us and our customers who use the harness to build and run their own agents. Improve the harness, and every surface it touches benefits.
### **Born in GitHub, battle-tested at enterprise scale**
The Copilot harness was built by the same GitHub engineers behind new technologies like our [multi-model orchestration](https://github.blog/ai-and-ml/github-copilot/project-hydrafusion-frontier-quality-via-multi-model-orchestration/). And those technologies are only possible because we built our own harness. Designed to be fast, token-efficient, and predictable, it was our original, high-quality harness. And it had the highest satisfaction among our most demanding audience: developers. That’s why we’re standardizing all of our Copilot experiences on the strongest foundation possible.
The harness is built to [perform across model families](https://x.com/ArtificialAnlys/status/2115477376981508419), with 20+ frontier models plus bring your own key, so you can use the right model for the right job. Instead of doing the work to keep up with and eval new models, lean on us to intelligently route work to the right model or models based on the task.
Our harness is flexible and configurable to any type of app, agent, or service. We recently [ported the harness runtime to Rust](https://github.blog/ai-and-ml/generative-ai/migrating-the-github-copilot-runtime-to-rust-using-copilot/), which has resulted in industry-leading performance and reductions in both session overhead and time-to-first token. And because the Copilot harness powers surfaces across Microsoft, it’s battle-tested and proven performant at the scale of millions of users and hundreds of real-world use cases.
A unified harness lets us ship faster, learn faster, and improve Copilot over time. And because the agents you build in Foundry can run on the same harness we’ve adopted across Microsoft, our continuous harness improvements will benefit your agents, too.
### **Try it yourself**
The [SDK is available to download](https://github.com/github/copilot-sdk), or you can run the harness in the cloud, all managed by Foundry. Our platforms remain open and interoperable, so you can continue to bring in any harness and framework you want, too.
With the Copilot harness, Microsoft runs the agent loop and sandbox, with governance and optimization built in from day one. You bring the instructions, skills, tools, and model choice.
This is an important step towards our vision of a stronger system around the model that preserves choice. It represents the best of GitHub, brought to the Microsoft platform and beyond.
We’ll keep pushing the harness forward, so you can focus on what only you can build.
The post [Introducing the Microsoft Copilot harness: Unlocking the potential of truly multi-model agentic work](https://blogs.microsoft.com/blog/2026/10/28/introducing-the-microsoft-copilot-harness-unlocking-the-potential-of-truly-multi-model-agentic-work/) appeared first on [The Official Microsoft Blog](https://blogs.microsoft.com/).