Schrondweiler Homelab:~$ ./
lang: en | fr
// DOCS
Glossary — Agent, Subagent, Skill, MCP
Portal — Public · Agents IA

← Agents IA

Documentation disponible en anglais uniquement.

#

Glossary — Agent, Subagent, Skill, MCP

These four terms get confused easily. Here's the distinction, with an analogy.

An analogy: a small IT team

Technical definitions

Agent (main). The session you drive directly — the only level with the full conversation history, and the one deciding how work is split.

Subagent. An isolated Claude instance spun up for one bounded task: isolated context (doesn't see prior conversation or files already read), a precise mission, only the tools it's been given, and it returns a final result rather than its intermediate steps. This lets work run in parallel and keeps the main session's context from being cluttered with details that only matter to the sub-task.

Skill. A file (often with supporting scripts or templates) that Claude loads automatically when a task matches its description — an extension of capability, usable by any agent, main or sub.

MCP (Model Context Protocol). The standardized protocol connecting Claude to external tools and services (databases, APIs, apps). It's what turns reasoning into an actual, real action on a system.

Why the distinction matters

A common misconception: assuming a "project manager" subagent can direct the others in real time like a human manager. It can't — each subagent is isolated and never talks directly to the others. That's exactly why the documentation-and-tracking agent in this setup only documents and tracks; the decision of who does what always stays with the main agent.

← Agents IA