What are Agent Skills? The folder-based standard for teaching AI agents tasks, now at the Linux Foundation
Agent Skills is an open format for teaching AI agents how to do specific tasks. A skill is a folder containing a SKILL.md instruction file plus optional scripts and resources; the agent keeps only each skill's name and one-line description in view and loads the rest when needed. Anthropic created it, opened it as a standard in December 2025 and donated it to the Linux Foundation's Agentic AI Foundation on October 22, 2026
The three lines
- Format — a folder: SKILL.md (name, description, instructions) plus scripts and files
- Mechanism — only descriptions load by default; the full skill loads on demand
- Status — donated to the Linux Foundation's AAIF on Oct 22, alongside MCP and AGENTS.md
Key questions
- Agent Skills vs MCP
- **MCP connects an agent to tools; a skill teaches it how to use them.** | | Agent Skills | MCP | |---|---|---| | Contains | Instructions, procedures, scripts | Tool and data connection protocol | | Form | Folder + SKILL.md | Client-server protocol | | Execution | Agent reads and follows | Server handles requests | | Analogy | Work manual | Power socket standard |
- How to write a SKILL.md
- **Name and description on top, instructions below.** | Part | Content | When loaded | |---|---|---| | Frontmatter (YAML) | name, description | Always (index) | | Body (Markdown) | Step-by-step guidance | When the skill applies | | Bundled files | Scripts, templates, references | When the body points to them |
- Which tools support Agent Skills
- **Anthropic says more than 70 agent platforms have adopted it.** | Category | Examples | |---|---| | Assistants | Claude, ChatGPT | | Coding tools | Cursor, GitHub Copilot, VS Code, Codex | | Steward | Linux Foundation AAIF (from Oct 22, 2026) |
The most tedious part of delegating to an AI agent is repeating the same instructions. "Use this report template, pull numbers from this sheet, run this check at the end." Agent Skills packages those instructions in a folder that the agent opens only when the task calls for it.
Anthropic created the format, opened it as a standard in December 2025, and on October 22, 2026 donated it to the Linux Foundation's Agentic AI Foundation (AAIF), which already hosts Anthropic's Model Context Protocol (MCP), AGENTS.md and Block's goose. A neutral foundation, not a single company, now stewards it.
1. Anatomy of a skill
A skill is a folder. Its core is SKILL.md.
| Part | Format | Example ("quarterly report" skill) |
|---|---|---|
| Frontmatter | YAML: name, description | description: Use when building a quarterly report from the finance sheet |
| Body | Markdown instructions | 1) Read the revenue tab 2) Fill the template 3) Run the totals check |
| Scripts | Python, shell, etc. | check_totals.py |
| Resources | Templates, reference docs | template.docx, style-guide.md |
Progressive disclosure
At start-up, the agent reads only each skill's name and description. When a task matches, it reads the body; when the body points to a file, it opens that file.
| Stage | What the agent reads | Context cost |
|---|---|---|
| Idle | Names and descriptions | Tens of tokens per skill |
| Task matches | SKILL.md body | Hundreds to thousands |
| Detail steps | Scripts and references | Only what's needed |
That is why dozens of skills can be installed without filling the context window.
2. How it differs from MCP and AGENTS.md
| Standard | One-line definition | Contains | Analogy |
|---|---|---|---|
| Agent Skills | Packaged know-how | Instructions, procedures, scripts | Work manual |
| MCP | Tool connection protocol | How to reach external systems | Power socket |
| AGENTS.md | Repository guide | Project rules and commands | Front-door sign |
| A2A | Agent-to-agent protocol | Task delegation and replies | Office messenger |
If MCP connects an agent to your database, a skill teaches it to "query it in this order and format the result this way."
3. What changed on October 22
| Item | Detail |
|---|---|
| Announcement | October 22, 2026, AGNTCon + MCPCon, San Jose |
| Joined alongside | OpenAI's Open Responses (one API across LLM providers), Hugging Face's OpenEnv (RL environments), Alibaba's OpenSandbox (agent sandboxes) |
| AAIF projects | 3 at launch in December 2025 → 10 |
| What Anthropic transferred | Specification, GitHub org, logo, agentskills.io domain |
| Adoption (Anthropic's figures) | 70+ agent platforms, 280 million installs |
MCP co-creator David Soria Parra said the goal is to let "anyone building with skills" shape what comes next. Rivals OpenAI and Alibaba placed their own standards in the same foundation on the same day — a push to keep agent tooling from fragmenting by vendor.
4. Uses and cautions
| Use | What goes in the skill |
|---|---|
| Company document style | Brand rules, formats, check scripts |
| Data analysis | Common queries, chart conventions |
| Code review | Team checklist, test commands |
| Customer support | Response principles, banned phrases, escalation rules |
- Scripts run. Read the scripts in any third-party skill before installing; skills can become a supply-chain attack vector, like npm packages.
- Prompt injection. The body is guidance the agent follows; a malicious line has the same effect as an injected prompt.
- The description is the trigger. Vague descriptions mean skills fire at the wrong time or not at all. Say precisely when to use it.
5. Quick answers
| Question | Answer |
|---|---|
| Do I need to code? | A guidance-only skill is a single Markdown file |
| Is it Claude-only? | No. It is an open standard, stewarded by the Linux Foundation since Oct 22 |
| Is it free? | The format is a free, open spec; running agents costs whatever the platform charges |
| Is it a plugin? | Similar, but skills center on instructions and load only when needed |