AI skills are moving beyond a single service: Agent Skills moves to AAIF

News: four projects under joint management

On October 22, the Agentic AI Foundation announced the joining of four projects: Anthropic's Agent Skills, OpenAI's Open Responses, Hugging Face's OpenEnv, and Alibaba's OpenSandbox. the foundation's announcement, all of them will operate under open governance without being tied to a single provider.

Agent Skills packages instructions, resources, and scripts for agent reuse. Open Responses concerns a shared interface for working with models, OpenEnv focuses on training and evaluation environments, and OpenSandbox on execution environments. These are different parts of a system, not four new chatbots.

What this changes for the average user

Our reading of the news: value is gradually shifting from a single successful prompt to one's own repeatable procedure. For example, rules for checking Ukrainian text can be gathered along with examples of the correct result. Such a procedure is easier to discuss, correct, and transfer than a lost fragment in a long conversation.

But an open format does not guarantee the same result in every program. You need to separately verify format support, available tools, and the specific agent's permissions. Changing the governance of a standard does not automatically install these skills on your computer.

Skill example for a Ukrainian AI media outlet

Here is a fictional editorial procedure: obtain the announcement text; write out verifiable statements; separate company claims from independent test results; save the event date; prepare a headline in Ukrainian; output a list of open questions.

It is worth adding a short example to such a package: the headline "AI has replaced programmers" is rejected if the source only describes an experiment on a single task. This is a specific quality criterion. The phrase "write coolly" does not provide one.

Before installing someone else's skill

• Read the instructions and see which files and commands they use.
• Check where additional scripts are downloaded from and whether they require network access.
• Start with a copy of training data without keys or client documents.
• Save the package version and the expected result: repeat the check after an update.

These are our recommendations for working, not a promise that any package marked with an open standard is safe. Trust in a format and trust in a specific author are separate decisions.

What you can do right now

Describe one of your own repeatable tasks: input data, steps, forbidden actions, and readiness criteria. Even without installing new packages, you will get a clear procedure for an assistant. Then test it on two different examples and fix the places where you had to guess.

To get started, we have a project context template and a an agent access map. The news is interesting precisely as a move toward portability; the functionality of a specific package still needs to be proven.

Sources and method

• AAIF — announcement of four new projects, 22.10.2026

Facts about the services have been verified against official materials. Examples, prompts, and the workflow were prepared by the editorial team. This is not a comparative test of model accuracy or speed. How we work.