On October 22, 2026, Anthropic announced that it was donating Agent Skills to the Agentic AI Foundation, a project under the Linux Foundation. Anthropic says more than 70 agent platforms had integrated the format, including ChatGPT, Cursor, GitHub Copilot and VS Code. The useful lesson for an AI learner is not the adoption count alone: it is the difference between a portable set of instructions and the software that carries those instructions out. The date, adoption figures and donation details come from Anthropic's announcement.
A skill is a folder, not a new model
The Agent Skills overview defines the core as a folder containing a SKILL.md file. That file includes a name, a description and instructions for a task. Optional files can include executable scripts, reference documents and templates. The format packages procedural knowledge: how to do a particular job, rather than just a description of the subject.
Imagine a practice skill for checking a draft explanation against a vendor document. Its instructions might identify the source, describe the checks and specify the result format. A reference folder could hold the checking rubric, while a script could perform a repeatable text check. This is an illustrative exercise, not a claim that such a checker has been built or that its conclusions would be correct.
Discovery is different from execution
The overview describes three loading stages. During discovery, the agent loads a skill's name and description. When the task matches, activation brings the full instructions into the agent's working context. During execution, the agent follows those instructions and may load supporting files or execute bundled code. The documentation calls this progressive disclosure: detailed material is loaded when needed, rather than loading every available skill in full at startup.
That distinction gives you a useful way to inspect a demonstration. Seeing a skill's name in a list shows discovery. Seeing the instructions loaded shows activation. Neither observation, by itself, demonstrates that a script ran successfully or that the resulting answer passed a check. For a small lab, record evidence for each stage separately instead of treating installation as proof of task completion.
What the foundation donation changes
Anthropic's announcement names the assets being donated: the Agent Skills name and logo, the GitHub organization and the agentskills.io domain. Engineers from several companies will maintain the standard as individuals. Using or contributing to it does not require membership in the foundation.
The announcement also describes versioned releases, a formal proposal process and security guidance as a proposed starting roadmap. Do not turn that wording into a claim that every part of the roadmap is already implemented. The donation concerns stewardship of the format; it is not a published audit of every skill someone might install.
Test portability without assuming identical behavior
The overview describes reuse across skills-compatible agents. For a learner, the practical exercise is to separate the reusable package from its execution environment. Start with a harmless task, inspect the instructions and any bundled scripts, then check whether each chosen agent can discover the skill, load its references and perform the requested steps. Compare the outputs against the same rubric.
Write down any environment assumptions rather than hiding them inside the word portable. A shared format is a starting point for that comparison, not evidence that two agents produced equally accurate work. This is a useful reading and testing exercise for AI learners, not a claim that the newly donated standard has been added to a particular certification exam.
Browse the Nex Arc study guides