How to set up Claude agents: all 7 types, step by step in Claude Code

Not technical? Start here. Install Claude Code (the first step under “Get set up” below), make a folder, start Claude in it, and paste the prompt. It walks you through all seven, one at a time, on made-up leads, and stops for your yes before anything costs money, reaches a real customer or deletes something.

This is the manual way: every file and every command. Claude follows it too. Here's how the seven fit together on one job, an incoming lead:

Words you'll see

• Agent: Claude working through a job on its own, step by step, using tools, instead of answering one question.
• MCP (Model Context Protocol): an open standard for plugging AI apps into your other apps. Its own site calls it a USB-C port for AI.
• Skill: a folder with a SKILL.md file: written instructions Claude loads when a job matches the description.
• Subagent: a helper Claude hands a task to. It works in its own fresh memory with its own instructions and tools, and only its summary comes back.
• Frontmatter: the few lines between --- marks at the top of a skill or subagent file: its name, what it's for, which tools it gets.
• Terminal: the text window where you type commands (Terminal on a Mac, PowerShell on Windows).

A tool agent is just Claude with a plug into your other apps, so it can look things up or take an action there. The plug is an MCP server. Lots of apps publish one; you add it with one command and log in.

A skill is your process, written down once, in a file Claude reads whenever the job matches. Claude only keeps the skill's name and description in mind until it's needed, so you can have a lot of them. The description has to say what it does and when to use it, because that's what Claude matches your request against.

Types 3 through 7 all run on subagents. A subagent is a text file in .claude/agents/ with a name, a description of when to use it, the tools it's allowed, and its instructions. Make these three once and you'll use them for the router, the chain, the parallel run, the boss and the checker. You can also just ask Claude to write a subagent for you; it writes the same kind of file.

This is type 5, and it's the right first test: it only reads, it writes nothing, and it uses the made-up leads. It runs on your Claude plan's normal usage, nothing extra.

Good: lead-1 comes back NEW JOB, lead-2 VENDOR, lead-3 CUSTOMER, each with a one-line reason quoting the message. In the transcript you'll see a row starting lead-router(, which is Claude handing it off. Bad: Claude answers without the subagent (restart, see above) or a label is off (make the label definitions in the file sharper; Anthropic's note is that routing works when the sorting is accurate).

A chain is one agent finishing and handing its result to the next. Claude Code's docs put it simply: ask Claude to use subagents in sequence, and it passes what matters from one to the next. Anthropic's tip is to put a check between the steps, so a bad first step doesn't roll downhill.

When the pieces don't depend on each other, run them at the same time and put the results together at the end. Each subagent works in its own memory, and Claude combines what they send back.

You built it in the first test. The router is where I'd start on inbound leads, because sorting is the boring part that eats the morning: what's a real job, what's a customer who needs a call back, and what's somebody trying to sell you leads. The router only labels. What happens next (a draft, a flag for the office) is a separate agent or a person.

The difference from parallel: with parallel, you decide the pieces. With an orchestrator, the boss looks at the job and decides the pieces itself, because you can't know them up front. In Claude Code the boss is your main Claude session; the workers are your subagents. Make it show you the plan first.

Subagents can hand work to subagents of their own, up to three layers below your main chat by default. For a small business that's more than you'll need.

This is the one I'd build first. One agent does the work and a different one grades it against a checklist, and the fixes go back to the first one. Anthropic says it works best when the grading rules are clear, so write your checklist like you'd write one for a new hire. The checker you made has no edit tools, so it can't quietly fix its own homework either.

Good: a PASS, or a FAIL with numbered fixes and then a better draft. Bad: it passes everything. Make the checklist stricter and more specific (word counts, banned words, the one question it must ask).

Subagent work counts toward the same usage limits as your main chat, so parallel runs and long fix loops use them up faster. That's why the loop above stops after two rounds.