Claude Code Sub-Agents vs Skills: The Hidden Power Features Nobody Explains

Claude Code has two features that change everything. Most developers use neither correctly.

Sub-Agents and Skills. They sound similar. They're completely different.

One spawns workers. One runs templates. Use the wrong one? You're wasting Claude's potential.

Let me show you exactly when to use each.

Sub-Agents are like hiring temporary contractors.

When you're working with Claude Code and it hits a complex task, it can spawn specialized sub-agents:

Types of Sub-Agents:

Explore Agent: Fast codebase exploration, finding files, understanding structure

Plan Agent: Designing implementation strategies, architectural decisions

Bash Agent: Running commands, git operations, system tasks

General Purpose Agent: Complex multi-step research tasks

Watch this. I ask Claude to 'understand how authentication works in this codebase.'

Claude spawns an Explore agent. It searches files, reads code, maps the auth flow - all while the main Claude waits.

Key Properties: Run in parallel with main conversation, Have their own context and tools, Report back when done, Can be run in background

I can spawn multiple sub-agents simultaneously. One explores auth, another checks the database schema, another reviews the API routes.

That's parallel AI workers. On your codebase. For $20/month.

Skills are completely different. Think of them as saved macros or templates.

You invoke them with slash commands: /commit, /review-pr, /pdf

How Skills Work: You type a slash command, Claude loads a pre-defined prompt template, The template executes with your context, You get consistent, repeatable output

Claude Code comes with built-in skills, but the real power is custom skills.

Here's one I built: /security-review

When invoked: Scan all changed files, Check for OWASP Top 10 vulnerabilities, Verify input validation, Check authentication flows, Output report with severity ratings

Every time I run /security-review, I get the same thorough process. No prompt engineering. No forgetting steps.

Key Properties: Invoked explicitly with /command, Run in the main conversation context, Template-based, consistent output, Can be shared across projects

Feature comparison: Invocation - Sub-Agents: Automatic or Task tool, Skills: /slash command

Execution - Sub-Agents: Parallel possible, Skills: Sequential

Context - Sub-Agents: Separate context, Skills: Main conversation

Best For - Sub-Agents: Exploration, research, Skills: Repeated workflows

Customization - Sub-Agents: Tool-based, Skills: Template-based

Persistence - Sub-Agents: Temporary, Skills: Saved permanently

Use Sub-Agents When: Exploring unfamiliar codebases, Need parallel processing, Complex research tasks, Want background execution

Use Skills When: Repeated, standardized tasks, Team-wide consistency, Complex multi-step workflows you run often, Want one-command execution

Here's the mental model:

Sub-Agents = 'Go figure this out'

Skills = 'Do this exact thing'

Both powerful. Different superpowers.

The real power? Combining both.

I built a skill called /full-feature that:

Spawns a Plan agent to design the implementation

Spawns an Explore agent to find relevant existing code

Waits for both to complete

Generates implementation based on their findings

Spawns a separate agent to write tests

One command. Multiple parallel agents. Consistent workflow.

What used to take me 2 hours of back-and-forth now takes 15 minutes.

Advanced Pattern: Agent Chains

Skills can invoke sub-agents. Sub-agents can trigger other workflows.

/implement-feature -> Plan Agent (architecture) -> Explore Agent (existing patterns) -> Main Claude (implementation) -> Bash Agent (run tests) -> Review output

This is agentic coding. AI orchestrating AI.

Example 1: Code Review Skill with Sub-Agents

/code-review - Spawns Explore agent to understand changed files' dependencies, Runs security checks, Checks for code duplication, Outputs structured review

Example 2: Research Task with Multiple Sub-Agents

'Help me understand how competitors handle rate limiting'

Spawns 3 Explore agents in parallel, Each searches different parts of codebase, Combines findings into comprehensive report

Example 3: Custom Deployment Skill

/deploy-staging - Runs tests via Bash agent, Builds project, Deploys to staging, Runs smoke tests, Reports results

One command. Full deployment workflow. Every time.

For Sub-Agents: Use 'run_in_background' for long tasks, Spawn multiple agents for parallel research, Choose the right agent type (Explore for search, Plan for design), Don't over-spawn - each agent costs tokens

For Skills: Build skills for tasks you do more than 3 times, Keep skills focused - one purpose per skill, Use clear /command names, Document what each skill does

General: Sub-agents for discovery, Skills for execution, Combine them for complex workflows, Monitor token usage - agents add up

I've put together a complete guide to Claude Code's advanced features at End of Coding.

Sub-agent patterns. Skill templates. Workflow examples.

Link in description.

Most developers use Claude Code like a fancy autocomplete.

Sub-Agents and Skills turn it into an AI workforce.

The question isn't whether to use them. It's which combination fits your workflow.