AAIF takes over four projects including Anthropic's Agent Skills

The Agentic AI Foundation (AAIF), under the Linux Foundation, announced on October 22, 2026 that it has received donations of four projects, including Anthropic's Agent Skills, which gives AI agents specialized knowledge and work procedures. The other donated projects are OpenAI's Open Responses, Hugging Face's OpenEnv, and Alibaba's OpenSandbox.

AAIF is an organization that supports the development of standard specifications and foundational technologies for AI agents from a vendor-neutral standpoint. At its founding in December 2025, its projects numbered three—MCP, goose, and AGENTS.md—but with the addition of these four, the total is now 10.

Agent Skills' operations transferred to AAIF

Agent Skills is a common format for reusing task-specific knowledge and work procedures across compatible agents. Procedures accumulated by a team and methods for checking results can be shared as skills and used across different compatible agents.

Each skill is created by bundling a SKILL.md file containing instructions together with any necessary scripts and reference materials into a single folder. Because agents load information in stages as needed, the amount of information loaded initially can be kept small even when many skills are available.

Anthropic released Agent Skills as an open standard in December 2025. According to the company, more than 70 platforms, including ChatGPT, Cursor, GitHub Copilot, and VS Code, have adopted it since its release.

With this donation, operation of the standard moves to AAIF. According to Anthropic's announcement the same day, the donation includes the Agent Skills name and logo, the GitHub organization account "agentskills," and the agentskills.io domain.

According to AAIF, a Technical Steering Committee, in which core maintainers from multiple organizations participate in a personal capacity, will set future development direction. Anthropic's announcement states that engineers from Anthropic, Google, OpenAI, and others will be involved in maintaining and updating the specification, with Anthropic's David Soria Parra and Mahesh Murag serving as lead maintainers.

As future initiatives, Anthropic proposes versioned releases, the establishment of a formal process for proposing specification changes, and the development of security guidelines for skill creators and adopters. In developing the specification, the policy is to emphasize community input and transparency in decision-making, as with MCP.

The other three donated projects

Open Responses: A common specification for model calls

Open Responses, donated by OpenAI, is an open specification based on the company's Responses API for calling language models from multiple providers in a common format. Because model APIs differ from provider to provider in the format of requests and responses, applications that support multiple models need processing to absorb those differences.

Open Responses standardizes the data structures for requests and responses, streaming for receiving generated results incrementally, the format for tool calls, and more. Model-providing services that implement this specification, as well as gateways that mediate connections to multiple models, can send requests using common code. This reduces the burden of building separate call-handling code for each provider.

What Open Responses changes for AI developers - AAIF

OpenEnv: Create and share environments for reinforcement learning

OpenEnv, donated by Hugging Face, is a framework for creating and sharing environments used for reinforcement learning of agents. Environment creators define tasks such as games, browser operations, and coding, along with methods for evaluating how well they are accomplished. The created environment receives an agent's actions and returns information observable after the action as well as numerical rewards that serve as learning signals.

OpenEnv handles task initiation, action submission, and result reception through a common API, making it easier to connect these environments to different training systems. The connected training system uses the information returned by the environment to train models. Environments are distributed as Docker images and can run locally, in the cloud, or on Hugging Face Spaces.

Training agents for real-world tasks with OpenEnv - AAIF

OpenSandbox: Isolated environments where agents work

OpenSandbox, donated by Alibaba, is a platform for building and operating isolated environments for AI agents on user-managed local or cloud infrastructure. Within these environments, agents can execute code, manipulate files, and operate browsers and desktops.

It supports Docker and Kubernetes, and handles environment creation and deletion as well as operations within environments through a common API. For example, when a coding agent fetches a repository and runs tests, it can create a working environment and delete it afterward. Because the same API can be used even if the execution infrastructure changes, the burden of building separate processing on the agent side is reduced.