
SubAgent vs Skills: Two Extension Methods for AI Agents
One-sentence explanation of the difference between SubAgent and Skills: An AI Agent is like an operating system, Skills are like applications, and a SubAgent is also an operating system, but pre-installed with apps and running like a virtual machine with an independent context.
Skills: Teach the Agent new skills
Skills are a "manual of operations."
You hire a smart assistant, but they know nothing about your business. Every time you assign a task, you have to explain the process, precautions, and tool usage from scratch. Skills turn this knowledge into documents that the Agent can refer to when needed, saving you from repeating yourself.
For example, I wrote a Skill called "Illustrate an article," which tells the Agent:
When the Agent starts, it only remembers "there is an illustration Skill," taking about 100 tokens. It reads the full details only when actually needed. This is called "progressive loading," which means loading many Skills won't blow up the context.

SubAgent: Spin up a virtual machine to work
A SubAgent is an independently running "expert assistant."
It has its own context window, equivalent to its own "working memory." Once the task is done, it only returns the result; the intermediate process does not pollute the main Agent's thinking space.
You are a project manager with three experts under you. You don't need to watch their every step; just tell them the task goal and wait for their reports.

A real case: Illustrating an article
Suppose I finish writing an article and want to create three sets of illustration schemes in different styles for the user to choose from.
If only using Skills:
Main Agent analyzes the article → designs three schemes → calls the illustration Skill to draw the first set → draws the second set → draws the third set
Problems arise:
If using SubAgent + Skills:
Main Agent analyzes the article → designs three schemes → starts three SubAgents simultaneously
Each SubAgent:
On the main Agent's side? The context is clean, with only three result links. The user can read the corresponding URL to view any scheme.

When to use which?
Use Skills:
Use SubAgent:
One-sentence summary
Skills are applications installed in the main system, called on demand; SubAgents are virtual machines that run independently and return results. Use applications for simple tasks, and spin up virtual machines for complex tasks.