Agent Skills: Anthropic launches skills for agents
2025-10-16 · Publisher : Anthropic · Domain : Artificial Intelligence
Anthropic introduces Agent Skills: folders of instructions, scripts and resources, described by a SKILL.md file, that Claude loads only when a task needs them. The format becomes an open standard on December 18, 2025.
What changes
- A skill is just a folder: a SKILL.md file, instructions, scripts and resources.
- Claude first reads only a short description and loads the full content only when the task calls for it.
- Available in the Claude apps (Pro, Max, Team, Enterprise), in Claude Code and through the API.
- On December 18, 2025, Anthropic publishes the specification as an open standard at agentskills.io.
- A team's know-how becomes a reusable, versioned asset that can be shared across agents.
On October 16, 2025, Anthropic introduces Agent Skills, a simple way to teach an agent specialized procedures: producing a document in a company's branding, applying an analysis method, following an internal process. Instead of repeating long instructions in every conversation, they are stored in folders the agent consults when needed. ### What happened - **The format**: each skill is a folder containing a SKILL.md file (name, description, instructions) and, where needed, scripts and resources. Anthropic describes them as "composable", "portable" and "efficient". - **Progressive disclosure**: the agent first reads only each skill's name and description, then loads the full instructions and files only when the task requires them. Many skills can thus be installed without saturating the context. - **Availability**: Claude apps for the Pro, Max, Team and Enterprise plans, Claude Code through a plugin marketplace, the API through a dedicated endpoint. A "skill-creator" skill helps write new ones. - **Open standard**: on December 18, 2025, Anthropic publishes the specification at agentskills.io for other platforms to adopt, and adds organization-wide management and a directory of partner skills. ### Why it matters After MCP, which connects agents to tools and data, skills standardize the transfer of procedural know-how. The same folder can serve several agents and be versioned like code. For an organization, it is a concrete way to capture its methods and spread them to all its agents. ### Caveats A skill can contain executable scripts: installing a skill of unknown origin carries a security risk comparable to that of a software dependency, and calls for review or scanning before use. ### Key takeaways Agent Skills turn business know-how into a shareable software component; opening them as a standard makes them a common building block of the agent ecosystem.
Links
Tags : IA, Agents, Standards ouverts, Anthropic, Claude