Skills and reusable setup

A skill is a named instruction package rendered beside the agent's AGENTS.md. Its description tells the model when to load it.

{
  "name": "release-qa",
  "description": "Use when checking release notes before publication.",
  "body": "Check every version, link, migration step, and compatibility claim against source.",
  "files": [
    {
      "path": "checklists/release.md",
      "content": "- Version is correct\n- Links resolve\n- Breaking changes are explicit\n"
    }
  ]
}

Constraints

• name uses lowercase letters, digits, and single hyphens, up to 64 characters.
• description is 1 to 1,024 characters.
• body is up to 50,000 characters.
• File paths are relative POSIX paths; absolute paths, backslashes, .., and root SKILL.md are rejected.
• Set both allow_executable_files: true and an executable file entry to ship a runnable helper. Sandbox policy must also permit execution.

Set disable_model_invocation: true to hide a skill from ordinary model selection and make it explicitly invokable only in supported harnesses.

Stored skills

A skills entry can be an @ag.embed reference to a stored skill workflow. Follow the skill workflow slug for the latest revision or pin a workflow revision version. Agenta resolves embeds before the runner starts.

Skills package instructions and files. They do not implicitly add an MCP server or connected application; configure those separately.