Mastering Skills, Tools, MCP, and Subagents in Anthropic’s Agent Course

Tools (底层能力)

Tools are the atomic operations that allow an agent to interact with external systems. They are always present in the context window.

Purpose – file reading, code execution, database access, etc.

Source – built‑in, custom, or loaded via MCP.

Tools = 底层能力 = 你有什么「工具」可以用

Skills (可重复的工作流)

Skills sit one level above Tools and describe *how* to use tools to accomplish a specific task. They are loaded only when needed (progressive disclosure).

Format – a standardized folder containing SKILL.md, command files, scripts, and resources.

Advantages – portable and reusable; write once, use many times.

Skills = 工作流程 = 用工具做什么「成品」

Core differences between Tools and Skills

Loading time – Tools are always in the context; Skills are loaded on demand.

Abstraction level – Tools are atomic operations; Skills compose those operations into workflows.

Complexity – Tools are simple; Skills can be arbitrarily complex.

Example – Tool: read a file; Skill: run a code‑review workflow.

Model Context Protocol (MCP) – 外部数据桥梁

MCP answers the question “what external data and tools does an agent need?” It connects the agent to external systems such as databases, Google Drive, and various APIs.

Provides – external tools and external resources.

Relation to Skills – Skills can leverage the tools and data supplied by MCP to complete tasks.

MCP = 外部连接器 = 获取 Agent 原本不知道的信息

Subagents (子代理) – Isolated execution units

Subagents are used when a task requires parallel execution or an independent context.

Features

Independent context windows

Fine‑grained permission control

Parallel task execution

Typical scenarios – code review, client‑interview analysis, multi‑document parallel processing.

Subagents = 分身术 = 让多个「小弟」并行干活

Collaboration patterns

MCP supplies low‑level capabilities (tools and data).

Skills define how to use those capabilities for a specific task.

The main agent can assign a Skill to a Subagent, e.g., a dedicated code‑review Subagent equipped with a code‑review Skill.

Practical case: Customer Insight Analyzer

The course presents a comprehensive example that combines all four components.

MCP provides access to external data sources (customer database, survey results).

Subagents run in parallel to analyze different data sources (interviews, surveys).

Skills ensure each step follows a standardized process.

Main Agent aggregates the results and outputs the final insight.

When to use each component

Tools – basic capabilities that the agent always needs (e.g., file I/O, code execution).

MCP – required when the agent must fetch external data or connect to external systems.

Skills – appropriate for repeatable, standardized workflows.

Subagents – useful for parallel processing or when an isolated context is needed.

Context‑window considerations

The context window is a shared resource across the whole agent system.

Subagents reduce pressure on the main context window by handling work in separate windows.

MCP loads external data only on demand.

Skills employ progressive disclosure, loading only when required.

Pre‑built Anthropic Skills

Skill Creator – a meta‑skill that helps create new Skills.

Excel/PowerPoint Skills – handle office documents.

Code Generation/Review Skills – support development workflows.