TL;DR：10 月 22 日，Linux 基金会旗下的 Agentic AI Foundation（AAIF）一口气收编四个项目：Anthropic 捐 Agent Skills（就是那个 SKILL.md 技能格式，70+ 平台、100 万份 Skills、近 2.8 亿次安装），OpenAI 捐 Open Responses（多供应商模型接口规范），Hugging Face 捐 OpenEnv（RL 训练环境框架），阿里捐 OpenSandbox（通用安全沙箱）。AAIF 成立才一年，托管项目从 3 个涨到 10 个——agent 技术栈的「获得技能、调用模型、训练评测、安全执行」四层，现在每层都有一个不由单一厂商说了算的家。冷思考也给你备好了：进基金会不等于从此中立，The Register 的评价是「厂商附带的机器被友好的基金会收养」。
资源在哪看
- Linux 基金会官方新闻稿：linuxfoundation.org
- Anthropic 官方博客：Donating Agent Skills to the AAIF
- AAIF 官方公告：aaif.io
- The Register 独立分析：theregister.com
📦 发生了什么：一场安排好的「四重奏」
时间地点都挑得讲究：10 月 22 日，AGNTCon + MCPCon North America 大会（圣何塞），AAIF 一次宣布四个新托管项目。这不是四家各自发新闻稿的巧合，而是打包好的联合动作——每家捐一件自己主导的基础设施，凑齐 agent 技术栈的四层：
至此 AAIF 托管项目从成立时的 3 个（MCP、goose、AGENTS.md）涨到 10 个，新增的还有此前的 A2A 协议、Agent Router、agentgateway。机器之心给了个精准的注脚：会员组织已达 305 家（此数字来自机器之心报道，官方稿未列会员数）——Agent 生态正在从各自造轮子，走向争夺公共标准。
🧩 四件东西分别管什么
Agent Skills 是四件里离开发者最近的。一个 Skill 就是一个文件夹：SKILL.md 写清楚名字、描述和操作步骤，旁边配上脚本和参考资料。设计上很克制——agent 启动时只读几十字的元数据，任务匹配了才读全文。Anthropic 自 2025 年 10 月发布以来攒下的家底全部随捐转移：名称、Logo、GitHub 组织、agentskills.io 域名，维护者是 MCP 联合创造者 David Soria Parra 和 Mahesh Murag，Anthropic、Google、OpenAI 的工程师以个人身份共同维护。
Open Responses 干的是「让不同模型说同一种接口语言」的活。各家 API 都能调工具、都能流式输出，但格式各有小性子，换模型就得改代码。Open Responses 基于 OpenAI 自家的 Responses API 定了一套共享 schema：请求输出描述一次，跨供应商运行。OpenAI 把自家 API 形状捐成行业标准，这步棋的意味值得细品。
OpenEnv 补的是训练侧。RL 训练 agent 得有环境反复练习，过去各团队重复搭轮子。OpenEnv 把环境做成容器化服务，一个接口同时对接 agent 协议、MCP 和训练库熟悉的 Gym。背后的共建名单很长：Microsoft、Reflection、Meta-PyTorch、Modal、Prime Intellect 都在里面。
OpenSandbox 管安全执行。agent 一旦会跑代码、操作浏览器、碰文件系统，就不再是聊天机器人了。阿里这个项目提供统一沙箱 API、多语言 SDK、Docker 和 K8s 运行时，还定义了生命周期与执行 API 的沙箱协议。阿里负责研发基础设施的 Xiaobin Xu 说得直白：「不让它的未来被单一公司定义。」
🏛 AAIF 凭什么当「中立之家」
AAIF 不是临时搭的草台班子。2025 年 12 月由 Anthropic 联合 Block、OpenAI 创立，隶属 Linux 基金会，创始托管项目就是 MCP——对，就是那个已经成为事实标准的 Model Context Protocol。一年里 A2A、agentgateway 等相继入驻，加上这批凑够 10 个。
AAIF CTO Manik Surtani 的话可以说就是这次事件的注解：「一年前我们带着三个项目和一个赌注成立——agentic AI 需要一个中立之家才能繁荣。今天四家不同的公司选择把它们交给中立治理，这就是基金会的职责：给社区一个共同建设标准的地方，没有任何单一公司控制结果。」
🛠 对开发者意味着什么
最直接的是防碎片化。拿 SKILL.md 打个比方：你给 Claude Code 写的技能，Cursor、ChatGPT、Gemini CLI、VS Code 都认——这个「一次编写、处处运行」的默契，以前靠的是大家自觉兼容，现在有了治理背书。格式留在单一厂商手里，迟早长出私有分叉；进了基金会，兼容性是义务不是恩惠。做过技能生态的都知道，这从「方便」变成了「保障」。
其次是接口层的靠拢。9 月 Anthropic 接受 AGENTS.md 标准、10 月 OpenAI 捐出 Responses 规范——头部厂商在接口层互相认账的速度在加快。对被多供应商适配折磨的开发者，这是实打实的减负。
最后是下半层终于有人管了。此前的中立治理集中在协议层（MCP/A2A/AGENTS.md），这次 OpenEnv 和 OpenSandbox 把「训练环境」和「安全执行」也拉进来了。agent 要跑得快，更要跑得安全——沙箱标准化的价值，最近各家的「智能体越界」新闻已经反复证明了。
🧊 不过，先别急着喊「永久中立」
把话说全：进中立基金会 ≠ 项目从此没有立场。The Register 的标题阴阳得很到位——「厂商附带的 AI agent 技术被友好的、有资金的基金会收养」。四个项目全是创建厂商主导捐入，技术路线的事实标准依然握在原厂手里；「以个人身份」任职的维护者制度能去多少厂商化，得靠时间验证。另外，Anthropic 提议的版本化发布、正式提案流程、安全指引目前只是「提议」，没有时间表。
但这些不构成看衰的理由。MCP 走过同一条路——捐赠时也没人信它会中立，一年后它成了行业事实标准。对写 SKILL.md、调多模型 API、搭 agent 沙箱的普通开发者来说，标准进基金会永远是比锁在单一厂商手里的更好结局。剩下的悬念只有一个：这四件「公共基础设施」，各家会不会捐完之后又悄悄在自己的产品里长出私有超集——老问题，新舞台。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
TL; DR: On October 22, the Linux Foundation compiled four projects in a single breath: Anthropic Donation Agent Skill (the SKILL.md skill format, 70+ platform, 1 million Skills, nearly 280 million installation), Open AI Donation Open Responses (multi-supplier model interface code), Hugging Face Donation OpenEnv (RL Training Environment Framework), Ali Donation OpenSandbox (General Security Box). AAIF has only one year since its inception, and the number of hosting projects has increased from 3 to 10 - the four levels of "Access to Skills, Call Models, Training Assessments, Safe Implementation" in the aggent technology vault, and now each level has a house that is not determined by a single manufacturer. It's not the same thing that goes into a foundation that is neutral since then.
Where are the resources?
-Linux Foundation official press release: linuxfoundation.org
- Anthropic Official Blog: Making Age Skills to the AAIF
- AAIF Official Gazette: aaif.io
- The Register Independent Analysis: Othergister.com
What happened?
Time and place were chosen: on 22 October, AGNTCon + MCPCon North America Congress (San José), AAIF announced four new hosting projects at a time. This is not a coincidence for the four individual press releases, but rather a packaged joint operation — each contributing its own infrastructure to bring together the four levels of the technology vault:
The number of AAIF hosting projects rose from three (MCP, Goose, AGENTS.md) at the time of their establishment to 10, with the addition of the previous A2A agreements, Agent Router, angentgateway. The machine's heart gives a precise note: the membership organizations have reached 305 (the figure is from the heart of the machine, the number of members is not included in the official text) — the Agent ecology is moving from building its own wheels to competing for public standards.
What about four things?
Agent Skills is the closest of the four exiters. A Skyll is a folder: SKILL.md writes a clear name, description and operational step, accompanied by scripts and references. Designed to be very restrained — agent reads dozens of words of metadata at startup, and the task matches the full text. Anthropic has moved with donations since publish in October 2025: names, Logo, GitHub, agentskills.io domain names, maintained by the engineers of MCP co-founders David Soria Parra and Mahesh Murag, Anthropic, Google, Open AI in their personal capacity.
Open Reponses do the job of "leting different models speak the same interface language". All APIs can adjust tools and flow outputs, but the formats are small, and the models change the code. Open Responses based on Open AI's own Reponses API set a shared schema: Request output description once, run across suppliers. Open AI donated its own API shape to industry standards, and this is worth a piece.
OpenEnv complements the training side. RL training antgents need to be practiced in an environment where teams used to repeat wheels. OpenEnv uses the environment as a containerization service, an interface with Gym familiar with the agent protocol, MCP and training library. The list is long: Microsoft, Reflecting, Meta-PyTorch, Modal, Prime Intellec.
OpenSandbox pipe secure execution. Once you run the code, run the browser, touch the file system, you're no longer talking robot. The Ali project, which provides unified sandbox API, multilingual SDK, Docker and K8s, also defines the life cycle and implementation of API sandbox protocols. Ali, Xiaobin Xu, who is responsible for developing infrastructure, put it bluntly: "Don't let its future be defined by a single company."
AAIF, what makes you a neutral house?
AAIF is not an ad hoc herbal team. Founded in December 2025 by Anthropic United Block, Open AI, affiliated with the Linux Foundation, the founding Trust Project is MCP - Yes, the Model Context Protocol, which has become the norm. A2A, angentgateway, and so on over a year, plus 10 of them.
The words of AAIF CTO Manik Surtani can be described as a note to this event: "One year ago we set up three projects and a bet — an entity AI needs a neutral house to prosper. Today, four different companies choose to leave them to neutral governance, which is the role of the Foundation: to give communities a place where standards are built together, without any single company controlling the outcome.
What does it mean to developers?
The most immediate is debris protection. Take SKILL.md, for example: your skills in writing Claude Code, Cursor, ChatGPT, Gemini CLI, VS Code, are all known as the "one-time writing, one-stop running", which is based on a self-fulfilling endorsement. The format remains in the hands of a single manufacturer, and sooner or later a private fork will grow; in a foundation, compatibility is an obligation, not a favour. Those who have done the skill ecology know that this has gone from "convenient" to "safeguard."
Next is the convergence of interface layers. September Anthropic accepts the AGENTS.md standard, October Open AI donates the Reponses norm - the speed at which head manufacturers are recognizing each other at the interface level is accelerating. This represents a real reduction for developers suffering from multiple suppliers.
Finally, the lower half is finally in charge. The previous neutral governance was concentrated on the protocol level (MCP/A2A/AGENTS.md), and this time OpenEnv and OpenSandbox brought in the "training environment" and "safe execution". Agent has to run fast and safely — the value of the sandbox standardization, as has recently been repeatedly demonstrated in the news of AI agent crossing the border.
But don't say "permanent neutrality."
Let's get to the bottom of this: The Neutral Foundation Project has never held a position since. The Register's title is very bright - "Accompanyed with the manufacturer's AI agent technology is adopted by friendly, funded foundations". The four projects were the creation of vendor-led donations, and the actual criteria for the technical route remained in the hands of the original plant; the system of defenders " serving in their personal capacity " would have to be tested over time for the extent of commercialization. Moreover, the version of publish proposed by Anthropic, the formal proposal process and the security guidelines are now merely "proposals ", with no timetable.
However, these do not constitute grounds for failure. MCP went the same way -- no one believed it would be neutral at the time of the donation, and a year later it became the standard of fact for the industry. For ordinary developers who write SKILL.md, multi-model API, and put on an individual sandbox, SRF is always a better end than being locked in the hands of a single manufacturer. There's only one thing left: these four "public infrastructure" will be donated and the families will be quiet. In their products, they grow private supersets — old problems, new arenas.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.