Agent Creator Guide FAQs

Question 1

Which AI model is best for specialized agents?

Accepted Answer

The choice depends on complexity: use Haiku for high-volume simple tasks, Sonnet for balanced coding tasks, and Opus for critical reasoning or complex security decisions.

Question 2

What is a Claude Code agent?

Accepted Answer

An agent is a specialized Claude instance configured with a specific set of tools, a focused system prompt, and defined responsibilities such as code review or security auditing.

Question 3

How do I test if my agent is working correctly?

Accepted Answer

The guide recommends a three-step process: run a baseline test without the agent, verify auto-delegation triggers, and perform edge-case testing with ambiguous inputs.

Question 4

How does Claude decide when to use a subagent?

Accepted Answer

Claude uses the 'description' field in the agent's frontmatter to match task triggers and symptoms. Writing this in the third person with specific keywords is critical for auto-delegation.

Question 5

Why should I restrict tools for my subagents?

Accepted Answer

Restricting tools (e.g., giving a reviewer only 'Read' access) ensures more focused behavior, reduces token costs, and prevents the agent from making unintended changes outside its scope.

Agent Creator Guide

Key Features

Use Cases

Agent Creator Guide

Key Features

Use Cases