Linux Foundation AAIF adds four Agent projects: Anthropic, OpenAI, Hugging Face, and Alibaba each donate one

Today, the Agentic AI Foundation (AAIF) under the Linux Foundation welcomes four new hosted projects at once:

Anthropic contributes Agent Skills, OpenAI contributes Open Responses, Hugging Face contributes OpenEnv, and Alibaba contributes OpenSandbox.

They cover multiple key stages of AI Agents, from acquiring skills and calling models to training and executing tasks. These are also four existing projects, brought under the same neutral governance framework.

With this, the total number of projects hosted by AAIF has increased from 3 at its founding to 10. Less than a year after the foundation was established, the number of member organizations has also reached 305.

It looks like just an open-source project donation, but looking deeper, things are not that simple.

When AI Agents start to actually do work, what the industry competes over is not just model capabilities, but also the rules that Agents commonly follow among themselves.

AAIF's new projects

What exactly are the four projects useful for?

These four projects each address one of the four problems in the process of an Agent doing work.

Agent Skills: Let Agents learn to do things according to procedures. This is a project contributed by Anthropic.

A Skill can be understood as a reusable work instruction manual. It usually includes SKILL.md, as well as scripts, reference materials, and so on needed to execute specific tasks.

For example, a company can package code standards, data analysis processes, and document templates into Skills. After compatible Agents load them, they can complete tasks according to these rules without having to explain everything from scratch each time.

Anthropic says that since this open standard was made public in December 2025, more than 70 Agent platforms have integrated it, with a cumulative 1 million Skills published and nearly 280 million installs.

These are figures disclosed by Anthropic itself, not the results of an independent audit. But at least it shows that Skills are no longer just a feature inside Claude.

Open Responses: Let different models speak the same interface language as much as possible. This is a project contributed by OpenAI.

Although the APIs of different model providers can all handle requests, call tools, and return results, the specific formats and streaming output methods may differ. If developers want to switch models, they often still have to modify a round of code.

Open Responses attempts to provide a unified interface specification, so that developers write less adaptation code and reduce the cost of switching models.

OpenEnv: Let Agents' training environments be reusable. This is a project contributed by Hugging Face.

Training an Agent cannot just let the model figure things out on its own; it also needs an environment where it can repeatedly act and receive feedback. OpenEnv hopes to make these environments into an open framework that is easier to create, share, and connect to training systems, reducing the work of teams repeatedly building infrastructure.

OpenSandbox: Give Agents a safe space to take action. This is a project contributed by Alibaba.

Once an Agent starts running code, operating a browser, and accessing files, it is no longer just a chatbot. OpenSandbox provides a unified sandbox API, multilingual SDKs, and runtime methods such as Docker and Kubernetes, allowing Agents to execute tasks in isolated environments.

The four projects each handle one area, but put together, they exactly cover several key stages of an Agent, from acquiring capabilities and calling models to training and executing tasks.

After MCP, Anthropic has handed over Agent Skills as well

This time, Anthropic not only contributed the Agent Skills specification, but also handed over the name, Logo, GitHub organization, and the agentskills.io domain to AAIF.

Note that this is not handing over the entire Claude or Claude Code product, but placing an open standard that has already formed an ecosystem under foundation governance.

When a technology is used by only a few products, it is not a big problem for the original developer to maintain it. But if it gradually becomes a foundational standard adopted by multiple platforms, other vendors will care: who exactly decides the future rules? If a company's product strategy changes, will the standard change with it? Do other participants have the opportunity to influence its direction of development?

Handing the standard over to a foundation is an attempt to reduce such concerns.

Anthropic also proposed a follow-up work plan, including versioned releases, a formal proposal process, and security guidelines for Skill writers and installers. Engineers from companies such as Anthropic, Google, and OpenAI will also participate in maintenance in a personal capacity.

However, donation is only the first step. True neutrality is not automatically achieved by putting a project into a foundation, but by whether other participants can truly participate in maintenance, modify specifications, and make decisions.

How do industry insiders view it? Some had already guessed it, some raised key questions

Simon Willison, an independent developer and well-known AI technology blogger, anticipated as early as December 2025 that Agent Skills might eventually enter AAIF.

However, he also pointed out at the time that this specification still had some details that were not clear enough. For example, developers can extend certain fields themselves, but different Agents may have differences in support for some functions.

To put it plainly, although everyone can use the same set of formats, that does not mean the products they build will be fully compatible.

Donating to the foundation now can provide a better governance foundation for improving the rules later, but it cannot automatically solve these technical problems.

Paul Nashawaty, Chief Analyst at ECI Research, offered another interpretation from an industry perspective.

In his view, the willingness of four companies to put different infrastructure projects into the same foundation shows that the industry is beginning to realize that there is no need to duplicate construction everywhere at the underlying infrastructure layer. Everyone can cooperate on interfaces and tools, and then compete over higher-layer model capabilities, product experience, and commercial markets.

But there is another, more subtle issue here. Even after projects enter the foundation, the original developers may continue to have a great influence on technical evolution because they are familiar with the code and possess the main implementation experience.

Therefore, a change in a project's nominal ownership does not mean that actual influence will necessarily be dispersed.

Another noteworthy voice came from the LinkedIn comment section under a post by Angie Jones, an AAIF ecosystem participant.

Commenter Roman Medvedev raised a key question: after these projects are placed in the same foundation, can they really achieve interoperability? Or will they still each go their own way?

This is also where technical standardization is easily overlooked. A foundation can provide a common governance framework, but it cannot guarantee that different products will fully implement the same set of specifications, let alone that they will perform consistently in real tasks.

Look at whether the community can truly participate, whether decisions are transparent, whether projects can be continuously maintained, and whether there are real implementations from different vendors.

In other words, do not just look at how many companies have announced support; look at what they have actually contributed and whether they can use the standard together.

Open Responses is worth watching

Open Responses is based on OpenAI's Responses API. Its goal is to let developers use a unified format for requests, responses, and tool calls, reducing the cost of adapting between different models.

This sounds like a developer experience issue, but it is actually also about vendor lock-in. If an Agent application can only rely on one company's proprietary interface, then switching vendors may mean rewriting a large amount of code. Conversely, if interfaces become increasingly standardized, developers can more easily choose models based on price, performance, and task requirements.

It is worth noting that AWS has publicly supported Open Responses and participated in project development, with core maintainers. Amazon Bedrock also already supports the Responses API, covering models from multiple model vendors.

This shows that the relevant interfaces already have a foundation in real product applications. But a line must also be drawn: supporting the Responses API does not mean that all models have passed full compatibility verification under the Open Responses specification.

Whether Open Responses can become a true cross-vendor standard in the future will depend on implementations by other vendors and on whether the community can jointly decide how the specification evolves.

This is also an interesting industry phenomenon: AWS's support for open interfaces can both reduce users' migration costs and help developers use different models on Bedrock.

Open standards and commercial competition are not contradictory. Companies can hope for more general underlying interfaces while continuing to compete on cloud services, models, and product experience.

Don't rush to declare that the Linux of the Agent era has arrived

This donation does send a positive signal: more and more companies are beginning to be willing to adopt open collaboration for Agent infrastructure.

Looking back at AAIF's first three projects—Anthropic's MCP, OpenAI's AGENTS.md, and Block's goose—plus the later additions of A2A, Agent Router, agentgateway, and these four new projects, the foundation's technology footprint is expanding.

MCP is responsible for connecting tools and data, AGENTS.md provides project-level instructions, Agent Skills encapsulates reusable capabilities, Open Responses promotes model interface interoperability, OpenEnv targets training environments, and OpenSandbox is responsible for isolated execution.

Together they are gradually sketching the outline of Agent infrastructure. But two test questions still have no answers.

First: after opening up, is it really safe?

In a study published on October 21, VirusTotal analyzed 35,878 AI Agent skill samples submitted to its platform, of which 6,637 were judged malicious.

Even more alarming, among these malicious samples, 62% contained no malicious code; the attack instructions were written directly in the natural language content of the skill documentation.

This shows that the security issues of Agent Skills are not just about detecting and removing malicious programs, but also involve instruction review, permission management, and execution environment isolation.

Of course, these figures come from samples received by VirusTotal and cannot be used to directly estimate the malicious proportion of the entire skills ecosystem, let alone to conclude that open standards themselves are unsafe.

Second: can neutral governance really be implemented?

After this project donation, what is worth watching next is whether there are more maintainers beyond the original contributors, whether the technical proposal and decision-making process are transparent, and whether different vendors can prove through actual testing that their implementations are compatible with one another.

Membership reaching 305 organizations shows that participation is expanding, but more members does not equal more contributions, still less does it mean that decision-making power has been fully dispersed.

So what is truly worth tracking is how many products from different vendors adopt these standards six months or a year from now, how many developers take part in the evolution of the specifications, and whether developers can change a model, tool, or execution environment with less code rewriting and fewer pitfalls.

Donating a project is only the opening act. Whether competitors can actually use the same set of rules to get things done is the real key to this standards battle.

Reference links

This article comes from the WeChat public account "Machine Heart" (ID: almosthuman2014), author: Neon Neon, published by 36Kr with authorization.