Agent Skills & Subagents · lesson 3 of 4

Claude Agent SDK

The library Anthropic provides for building custom agents that share Claude Code's primitives — tool loop, permission system, hooks, subagents, skills — without writing them from scratch.

What you get for free

Multi-turn agent loop with tool use.
Permission middleware (allow/deny + user confirmation).
Hook system for lifecycle events.
Subagent delegation.
Skill loading.
Session management.
Streaming + output formats.

Languages

TypeScript / JavaScript (@anthropic-ai/agent-sdk).
Python (anthropic-agent-sdk).

Sketch (TypeScript)

import { Agent } from "@anthropic-ai/agent-sdk";

const agent = new Agent({
  model: "claude-sonnet-4-6",
  systemPrompt: "You are a build engineer for our monorepo...",
  tools: [readFileTool, runShellTool, githubMCP],
  permissions: {
    allow: ["Read(./**)", "Bash(pnpm:*)"],
    deny: ["Bash(rm:*)"],
  },
  hooks: {
    PostToolUse: [{ matcher: "Edit", command: "pnpm lint --fix" }],
  },
});

const result = await agent.run({
  prompt: "Bump deps to latest minor and re-run tests.",
});

When to use the Agent SDK vs vanilla API

Need	Use
Quick one-off tool-use script	Vanilla Messages API
Reusable agent with permissions, hooks, subagents	Agent SDK
Coding agent for your repo	Claude Code (it IS Agent SDK + ergonomics)
Production multi-tenant agent	Agent SDK + your infra

Custom subagents

Define subagent types as you would in Claude Code (.claude/agents/*.md). The SDK loads them automatically.

CCA-F angle

You're unlikely to see Agent SDK syntax on the exam, but you should know it exists and that Claude Code is built on top of it. Architecture questions about "how would you build a custom Claude-powered agent" expect you to recognize the SDK as the answer, not a from-scratch tool loop.