Agent / Sub-agent / Skill / MCP — 4 Roles in 3 Lines
Answered in 3 lines
- Agent = the "person" who thinks and decides; Sub-agent = a specialist team inside an Agent
- Skill = an instruction sheet for "how to do" something; MCP = a "connection" to external systems
- They are not the same. They differ along four axes: who, what they know, what they connect to.
At-a-glance mapping
| What you want to do | Agent | Sub-agent | Skill | MCP |
|---|---|---|---|---|
| Talk with the user and understand the task | ✅ | ❌ | ❌ | ❌ |
| Orchestrate the entire job | ✅ | ❌ | ❌ | ❌ |
| Work on a specialty in an isolated context | ❌ | ✅ | ❌ | ❌ |
| Objective quality gate (review, validation) | ❌ | ✅ | ❌ | ❌ |
| Teach "how to create a PR" procedure | ❌ | ❌ | ✅ | ❌ |
| Convey coding conventions / domain knowledge | ❌ | ❌ | ✅ | ❌ |
| Call an external API | ❌ | ❌ | ❌ | ✅ |
| Query an internal database | ❌ | ❌ | ❌ | ✅ |
| Access the filesystem | ❌ | ❌ | ❌ | ✅ |
The Four in One Diagram
- Agent / Sub-agent = actors (subjects)
- Skill = knowledge and procedures the actor references
- MCP = the connection point through which the actor reaches the world
Common search questions, answered in 3 lines
Q: Difference between Skill and Sub-agent in one line?
A: Skill expands in the parent context; Sub-agent launches in an isolated context. The dividing question is whether intermediate tool calls flow into the parent. See Sub-agent vs Skills.
Q: What's different between MCP and Sub-agent?
A: MCP is a "connection"; Sub-agent is a "specialist." MCP is a server process (an opening to external systems); Sub-agent is a separate persona running inside Claude Code. The Sub-agent uses MCP from the inside.
Q: Which of the four should I build first?
A: Start with Skill. A single Markdown file is enough; the ROI is immediate. Next comes MCP (when external connection is required), and last Sub-agent (once isolated context becomes a hard requirement). Agent (main) is already provided by hosts like Claude Code.
Q: How are Agent and Sub-agent related?
A: Sub-agent is a form of Agent. Specifically, "a child Agent delegated by a parent and executed in an isolated context." It is general Agent / custom agent, narrowed by lifecycle attributes (Ephemeral / Spawned). See Agent Taxonomy.
Q: Do I have to use all four?
A: No. Use only what your task needs. For example: "teach coding conventions" — Skill only; "call an external API and show results" — MCP only; "exploratory codebase investigation" — Sub-agent only. All are fine.
Q: What is the typical composition pattern?
A: The 3-layer pattern: "Skill for procedure, Sub-agent for execution, MCP for connection." Example: translation workflow → Skill translation-workflow defines the procedure, Sub-agent translator executes as the specialist, MCP deepl-mcp calls the translation API.
Q: Where do meta-agents, Orchestrator, Swarm fit?
A: These are design patterns, not implementation units. Agent / Sub-agent / Skill / MCP are implementation units; Orchestrator-Worker and Swarm are architectural patterns that combine them. See Agent Taxonomy.
Q: When the same goal can be implemented as either Skill or Sub-agent, which?
A: Default to Skill. Reasons: lower startup cost, less host dependency. When promotion conditions (Skill → Sub-agent) emerge, migrate. Promotion signals: "parent context inflates," "parallelism needed," "objectivity required." See Sub-agent vs Skills / When to promote.
Decision flow (decide in 15 seconds)
Going deeper
| What you want to know | Page |
|---|---|
| MCP vs Skills, 3-line answer | MCP vs Skills FAQ |
| Detailed Skill vs Sub-agent selection | Sub-agent vs Skills |
| Using Sub-agents as quality gates | Using sub-agents as quality gates |
| Agent terminology organization (Orchestrator, Swarm, etc.) | Agent Taxonomy |
| Sub-agent basics | What is a Custom Sub-agent |
| MCP basics | What is MCP |
| Skill basics | What is Skills |
| Architecture overview | 03-architecture |
| Relation to the Memory layer | 08-memory-and-knowledge |
Next: MCP vs Skills FAQ