liveBook · Manning

Agent skill

As explained in Build a Multi-Agent System (from Scratch)

A skill is a reusable workflow packaged for discovery, activation, and execution by an LLM agent.

An agent skill is a reusable, documented workflow that an LLM agent can discover, activate, and execute. A skill defines the steps for a task; tools provide discrete functionality the agent can invoke along the way. The workflow is written once rather than repeated in every prompt. section 6.1.1 section 6.6

Maintaining a research list

A skill can instruct an agent to find a paper, categorize it, summarize it, and add it to a list. The agent may use web search, file-reading, and board tools while following the saved procedure. section 6.1.1

The Agent Skills open standard packages a workflow in a directory containing SKILL.md. Its YAML frontmatter requires a name and description; the Markdown body contains the procedure. Optional scripts/, references/, and assets/ directories hold executable code, supporting documentation, and other files. The workflow can reference resources at the steps where they are needed. section 6.1.2 section 6.1.1

Table 6.1 Fields that can be included in a skill’s frontmatter
Field
Required
Description
name
Yes
A unique identifier for the skill. Must be lowercase letters, numbers, and hyphens only. Cannot start or end with a hyphen. Maximum 64 characters.
description
Yes
Describes what the skill’s procedural workflow does and in what scenarios it should be applied. Maximum 1024 characters.
license
No
The name of the license that covers the skill (e.g. MIT) or the name of the license file that is bundled in the skills resources.
compatibility
No
Describes any environment constraints or requirements the skill has. Maximum 500 characters.
metadata
No
An arbitrary key-value map for additional properties not covered by the standard, like author or version.
allowed-tools
No
A space-delimited list of tools (each specified by name) the LLM agent can use to execute the skill.

Use the frontmatter field table to check required fields, naming constraints, and optional metadata. section 6.1.2

A skill-enabled agent uses a staged lifecycle:

  • Discover skills from the configured scope and disclose a catalog of their names, descriptions, and locations to the model. section 6.6 section 6.2.2
  • Activate a selected skill to disclose its complete information, including the workflow and available resources. The model can use UseSkillTool, or a user can request explicit activation with run_with_skill(). section 6.6 section 6.1.2
  • Execute the workflow, using tools and referenced resources as needed. section 6.1.1 section 6.1.2

A Skill implementation stores frontmatter, location, and scope. Its resources property finds bundled files; catalog() formats discovery information; and read_body() loads the workflow, raising errors for unreadable files, invalid frontmatter separation, or an empty body. section 6.2.2

Scope controls availability: user-scoped skills are accessible across projects, while project-scoped skills are limited to their project. In the illustrated framework, discovery occurs on each run, and scope selection is coarser than assigning individual tools. The explicit_only_skills parameter can hide specified skills from autonomous model activation. section 6.6 section 6.5

How Build a Multi-Agent System (from Scratch) covers it

How deep it goes
Works through bundle structure, skill representation, discovery, activation, execution, and scope-based access control.
Examples
Uses a research-list workflow and exercises involving an MCP tool, explicit activation, and resource directories.
What it assumes
Builds on existing LLM agent and MCP tool infrastructure while introducing skill specification and lifecycle.
What you can do after
Builds skill support from scratch and applies it to reusable workflows with model-selected or user-explicit activation.
How it connects
Connects skills to tool calling, Model Context Protocol, and multi-agent systems.

Also explained in

Compare how 6 books explain agent skill →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Build a Multi-Agent System (from Scratch). The section links point to the parts of the book it draws on.