Claude Skills vs Agents & Subagents: The Hierarchy

Understanding the Architecture via Analogy

To understand the difference, imagine a professional kitchen.

The Agent (The Head Chef):

• The main entity interacting with the user. It has general knowledge and coordinates the workflow. It decides what needs to be done.

The Skill (The Recipe):

• A specific set of instructions (e.g., "How to make a Soufflé"). The Chef reads the Skill to execute the task perfectly. A Skill is passive; it doesn't do anything until an Agent reads it.

The Subagent (The Sous Chef):

• If the Head Chef is too busy, they hire a Subagent (Sous Chef) solely to chop vegetables. This Subagent has a specific role, uses specific Skills (Knife Skills), and reports back to the main Agent when finished.

Technical Definitions & Roles

1.Claude Agent

• Definition: The primary LLM instance (e.g., Claude 3.7 Sonnet) running in a loop.
• Role: Holds the conversation context, maintains memory, and decides which tools to call.
• Lifespan: Persists throughout the entire user session.

2.Claude Agent Skill (SKILL.md)

• Definition: A static file containing procedural knowledge and executable scripts.
• Role: Provides the procedure. It teaches the Agent how to interact with a specific API, database, or codebase.
• Lifespan: Passive. Loaded into context only when needed.

3.Subagent (The fork Context)

• Definition: A new, isolated Agent session spawned by the main Agent.
• Role: Executes a complex, multi-step task (e.g., "Research this topic") without polluting the main Agent's context window.
• Lifespan: Temporary. Created for a task, destroyed after returning the result.

Skills vs. Subagents: Feature Breakdown

Decision Guide: Skill or Subagent?

Scenario A: "I need to format a JSON file."

• Solution: Use a Skill.
• Why: It's a deterministic, linear task. The Agent just needs the instructions (Skill) to do it.

Scenario B: "I need to browse the web and find the best library for my project."

• Solution: Spawn a Subagent.
• Why: This requires exploration, trial and error, and reading many pages. If the main Agent does this, it will fill up its context window with irrelevant web pages. A Subagent can do the messy work and just report back: "I found library X."

Scenario C: "I need to refactor 50 files."

• Solution: Verdent Parallel Subagents.
• Why: Don't do it sequentially. Spawn 50 Subagents, each equipped with the refactor-skill.

The Orchestrator: Verdent

Managing one Agent is easy. Managing a team of Agents and Subagents using hundreds of Skills requires an OS. Verdent is that OS.

• Skill Management: Verdent injects the right SKILL.md into the right Agent.
• Subagent Spawning: Verdent visualizes Subagents as "Parallel Tracks." You can see the main thread and 3 sub-threads working simultaneously.
• Context Isolation: Verdent ensures that the "messy thinking" of a Subagent doesn't leak into your main project context, keeping your workspace clean.