Agent Skills Linux Foundation: the format that teaches your AI your processes is now open
Agent Skills moves to the Linux Foundation: how a skill works, three business cases and the security number almost nobody mentions.
Agent Skills Linux Foundation is the October 22, 2026 story: Anthropic handed the skills format, folders with a SKILL.md file that teach your AI a process, to the Agentic AI Foundation, which lives inside the Linux Foundation. The news isn't the giveaway, it's who now controls the format: what you write today depends less on a single company.
What changed with Agent Skills and the Linux Foundation
Almost everyone will tell this as "Anthropic gave away its skills." That's not it. The news is who governs the standard. On October 22, in San Jose, the Agentic AI Foundation added four projects and reached ten. One is Agent Skills, contributed by Anthropic; another is Open Responses, from OpenAI.
Anthropic handed over the standard, the name and the logo, the GitHub organization, and even the domain, agentskills.io. Before, the format had an owner. Now it's governed by a technical committee of maintainers from several companies serving as individuals, with engineers from Anthropic, Google and OpenAI. Skills were born in October 2025 and the standard was opened that December.
How a skill works inside
A skill is a folder. Inside is a SKILL.md file with two pieces of data at the top: the name (lowercase with hyphens, matching the folder) and a description of up to 1,024 characters that says what it does and when to use it. Below come the instructions in plain text, and optionally folders with scripts, references and templates.
The key piece: the agent doesn't read everything, it loads in stages. First only the name and description of each skill (about 100 tokens). If your request matches, it opens the full SKILL.md. Only if needed does it open the supporting files. That's why you can have dozens of skills without filling up the agent's memory.
The walkthrough with a quoting skill
- Create a folder called agency-quotes; the folder name must match the name field.
- Inside, a SKILL.md with a description like "generates quotes with our rates and format; use it when someone asks for a proposal, a bid, or a quote."
- Put your rates in references, and a short program that adds up the totals in scripts.
- Install it in your tool and ask for a quote for a five-page website.
Where people get stuck the first time: the vague description. If it says "helps with quotes," it almost never triggers. To check the format, the project ships skills-ref validate.
The math: 250,000 versus 9,900 tokens
This is a WAiBOT example using the figures the specification recommends, not a measurement. With 50 skills at 5,000 tokens each, loading everything would be 250,000 tokens before you type a word. With staged loading, 50 descriptions at 100 tokens come to 5,000; when you use one, 49 times 100 plus 5,000 is 9,900. That's about 25 times less, and it's also why a bad description costs so much: if it doesn't trigger, it's as if the skill didn't exist.
Three business cases (WAiBOT analysis)
These cases are our own analysis, not published cases or market prices.
- A four-person marketing agency: one proposals skill with rates, style and structure works the same in Claude, Cursor or Copilot. When a rate goes up, they change one file and all four proposals come out with the new price.
- An accounting firm: the monthly close skill carries each client's checklist and a script that verifies the totals add up. The script does the math, not the model, and since it lives in Git there's a record of who changed which rule.
- A sector consultant: packages how to qualify a prospect, what to answer to each objection and how to follow up, and installs it in each client's tools. Anyone can copy the folder, but not the judgment of whoever wrote it.
Who it's not for, and how neutral it is
It isn't for people who only use AI for one-off questions with no process to repeat; for companies that don't allow an agent to run scripts or install third-party code; or for anyone whose tool doesn't support skills yet.
The "this is theater" objection is partly right: the word "neutral" comes from Anthropic's announcement, and one of the two maintainer leads, David Soria Parra, is from Anthropic. But the name, the logo, the GitHub account and the domain no longer belong to Anthropic, and that's hard to reverse. The objection would win if in six months there are no numbered versions and no formal proposal process, which today are only proposed.
What nobody tells you: the adoption number and security
The claim going around (more than 70 platforms, a million skills and almost 280 million installs) comes from Anthropic's announcement, with no independent audit and no definition of "install." Don't take it as a guarantee: test your skill in the two tools you actually use.
What really breaks is security. A skill inherits your agent's permissions: files, terminal, credentials. In February 2026 Snyk reviewed 3,984 skills from ClawHub and skills.sh: 36.82% had at least one flaw, 13.4% a critical one and 76 carried confirmed malicious code. It's one security company's report, on other catalogs, and the standard's security guidance is only proposed today.
How to apply it in your business?
- Write your skill with a lowercase name and a description in the words you'd actually use to ask for it.
- Keep the SKILL.md short (under 500 lines) and send the detail to references.
- Keep it in Git so you know who changed what.
- Before installing a third-party skill, read it all, scripts included, and distrust any that ask you to download password-protected files or run base64 text. Snyk recommends the open-source scanner mcp-scan.
- Test it in two different tools, and never put keys or passwords inside a skill.
To see the same format from Google's side, read the full breakdown of Gemini skills replacing Gems.
Agent Skills and the Linux Foundation: frequently asked questions
What is Agent Skills?
An open format: a folder with a SKILL.md file (name, description and instructions) and, optionally, scripts, references and templates, that teaches an AI agent a specific process.
What changed on October 22, 2026?
Anthropic donated the standard, the name, the logo, the GitHub organization and the agentskills.io domain to the Agentic AI Foundation, inside the Linux Foundation. A technical committee of maintainers from several companies governs it.
Will my Claude skills work in other tools?
The format is the same and Anthropic says more than 70 platforms use it, but that number is self-reported. The prudent move is to test your skill in the tools you use.
Is it safe to install third-party skills?
Not without reviewing them: a skill inherits your agent's permissions. Snyk's February 2026 report found at least one flaw in 36.82% of 3,984 skills reviewed.
Conclusion
With Agent Skills in the Linux Foundation, what you write today depends less on a single company, and the risk is now what you install, not where. If you write one skill well this week, you use it in all your tools; if you install a third-party one blind, you open the door to your terminal.