OpenAI 与四家竞争对手就 AI Agent 统一标准达成一致

英文标题：🔶 OpenAI and four rivals just agreed on one standard for AI agents

OpenAI、Amazon、Cursor、Microsoft 和 Vercel 等公司发布了开放标准 Agent Plugins，旨在让同一套 AI Agent 扩展跨不同产品运行。该标准将 Model Context Protocol（MCP）服务器与可复用的 Agent Skills 打包为统一插件格式，核心文件为位于根目录的 `plugin.json`。ChatGPT、Codex、Cursor、GitHub Copilot、Kiro 和 VS Code 已在发布时支持这一格式。标准刻意未规定插件市场、安装、权限、沙箱和信任机制，因此互操作性提升的同时，安全审核与平台控制权等关键问题仍留给各客户端解决。

打开原文相关 5 · 机会 1

中文

AI 竞赛过去两年主要围绕模型展开。本周，焦点转向了底层基础设施。在 GPT-5 一周年前夕，OpenAI 与四家竞争对手发布了开放标准 Agent Plugins，让单个 Agent 扩展能够在相互竞争的产品之间通用。

其主张是“构建一次，随处运行”。如今，每个 Agent 工具都要求不同的文件夹结构和配置方式。Agent Plugins 用一种统一的软件包格式取而代之。它打包了开发者已在使用的两类内容：第一类是 Model Context Protocol servers，用于将 Agent 连接到实时工具和数据；第二类是 Agent Skills，即可复用的指令集合。一个插件本质上就是一个文件夹，其根目录中包含一个小型 plugin.json 文件。

谁在推动这一标准

尽管 OpenAI 是宣传中的主角，但这并不是 OpenAI 的产品。 Vercel 发起了这项提案。此后，来自 Amazon、Cursor 开发商 Anysphere、GitHub、Microsoft、OpenAI 与 Vercel 的代表共同制定了 1.0 规范。指导委员会由五个成员构成：Amazon、Cursor、Microsoft、OpenAI 和 Vercel。该项目采用开放许可证，其支持者表示，没有任何一家公司的路线图能够单独决定项目方向。

OpenAI 仍将这次发布定位为自己的一个里程碑。根据 9to5Mac 的报道，GPT-5 将于 8 月 7 日迎来一周年，OpenAI 正借此在本周将视野投向单个模型之外。其 ChatGPT 和 Codex 应用在发布时即支持该格式，Cursor、GitHub Copilot、Kiro 和 VS Code 也同样支持。这一举措比 OpenAI 自家的 Codex plugins 更进一步，后者只能在 OpenAI 自身工具内部运行。

有意保持精简

该标准刻意保持狭窄的范围。它只定义插件如何打包、如何被发现，几乎不涉及其他内容。插件市场、安装、权限、沙箱和信任机制均由各客户端自行负责。这使得该格式更易于采用，但也让最棘手的问题悬而未决。判断一个插件能否安全运行，依然是每个客户端自己的职责；而在今年早些时候，伪造的 Agent Skills 曾绕过安全扫描器，这正是一个持续存在的现实担忧。

并非所有人都信服。开发者工具框架 SST 的构建者 Dax Raad 表示，他“非常反对”这一标准，并称它是“一个薄弱的标准”，其真正有用的部分最终仍会落入客户端专属扩展之中。

也有人对此表示欢迎。开发者布道师 Angie Jones 写道：“我们太需要这个了（We neeeeded this）。”她此前一直希望能有一种方式，将自己的 Skills 在所使用的不同工具之间携带和迁移。

更大的问题在于：将底层基础设施标准化会如何影响竞争。统一格式可能让小型开发者一次就触达所有主流 Agent。这是开放生态系统的论点，而且确实成立。但它也可能巩固少数已经拥有用户的客户端，因为标准会奖励那些用户本来就在使用的平台。

目前，底层基础设施已经达成共识。至于标准刻意留下的部分——插件市场与信任机制——围绕它们的竞争才刚刚开始。

English / 原文

The AI race spent two years being about models. This week it moved to the plumbing. On the eve of GPT-5’s first birthday, OpenAI and four rivals published Agent Plugins, an open standard that lets a single agent extension work across competing products.

The pitch is “build once, run anywhere”. Today, every agent tool expects a different folder layout and setup. Agent Plugins replaces that with one package format. It bundles two things developers already use. The first is Model Context Protocol servers, which connect an agent to live tools and data. The second is Agent Skills, reusable sets of instructions. A plugin is just a folder with a small plugin.json file at its root.

Who is behind it

Despite the OpenAI billing, this is not an OpenAI product. Vercel initiated the proposal. Representatives from Amazon, Cursor’s maker Anysphere, GitHub, Microsoft, OpenAI, and Vercel then shaped the 1.0 specification. The steering committee has five members: Amazon, Cursor, Microsoft, OpenAI, and Vercel. The project is openly licensed, and its backers say no single company’s roadmap sets its direction.

OpenAI still framed the launch as its own milestone. GPT-5 turns one on 7 August, and the company is marking the week by looking beyond individual models, per 9to5Mac. Its ChatGPT and Codex apps support the format at launch, alongside Cursor, GitHub Copilot, Kiro, and VS Code. The move goes further than OpenAI’s own Codex plugins, which work only inside its tools.

Deliberately small

The standard is narrow on purpose. It defines how a plugin is packaged and found, and little else. Marketplaces, installation, permissions, sandboxing, and trust all stay with each client. That keeps the format easy to adopt. It also leaves the hardest problems unsolved. Deciding whether a plugin is safe to run is still every client’s own job, a live worry after fake Agent Skills slipped past security scanners earlier this year.

Not everyone is convinced. Dax Raad, who builds the SST developer-tools framework, said he was “very much against” it, calling it “a thin standard” whose useful parts will end up in client-specific extensions anyway.

Others cheered. “We neeeeded this,” wrote developer advocate Angie Jones, who had wanted one way to carry her skills between the tools she uses.

The bigger question is what standardising the plumbing does to competition. A shared format could let a small developer reach every major agent at once. That is the open-ecosystem case, and it is a real one. It could also cement the handful of clients that already have the users, because a standard rewards whoever people already run.

For now the plumbing is agreed. The fight over the parts it leaves out, the marketplaces and the trust, has barely started.