The 7 kinds of Claude agents in plain English, each built in Claude Code: an app plug, a skill, a chain, parallel agents, a router, a boss and a checker.
Swipe or tap Next · checked on 2026-10-31
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.
Make the folder and start Claude in it (Mac, Linux, WSL)
mkdir -p ~/agent-shop/leads ~/agent-shop/drafts
cd ~/agent-shop
claudeThe setup prompt (copy all of it) · lines 1-1 of 24. Copy all copies the whole prompt; read it in full on the page.
Hi Claude. Please be my patient setup helper. I run a small business and I don't code: one step at a time, tell me exactly what to type, explain why, and wait for me to say "done" before you move on.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:
Got ten leads instead of one? Claude splits them across several writers at once (type 4, parallel) or plays the boss and decides the split itself (type 6, the orchestrator).
claude mcp addSKILL.md file in .claude/skills/.claude/agents/Pattern names from Anthropic's Building effective agents; Claude Code features from its docs (checked 2026-10-31).
SKILL.md file: written instructions Claude loads when a job matches the description.--- marks at the top of a skill or subagent file: its name, what it's for, which tools it gets.Anthropic's advice is to find “the simplest solution possible, and only increasing complexity when needed.” Most jobs don't need all seven. A router and a checker cover a lot of ground.
Claude Code needs Pro, Max, Ultra, Team, Enterprise or a Console account. The free plan doesn't include it. Pro is $20 a month billed monthly (checked 2026-10-31).
Open a terminal and run the line for your computer. When it finishes, open a new terminal window and check it.
Mac, Linux, WSL
curl -fsSL https://claude.ai/install.sh | bash
# open a NEW terminal window, then:
claude --versionWindows PowerShell
irm https://claude.ai/install.ps1 | iexThe first time you run claude, it opens your browser to log in to your Claude account.
Make the folder and start Claude
mkdir -p ~/agent-shop/leads ~/agent-shop/drafts ~/agent-shop/.claude/agents ~/agent-shop/.claude/skills/lead-reply
cd ~/agent-shop
claudeTest on fake ones first, so nothing real gets touched. Save each as its own file in leads/.
Example (made up): three leads · Part 1 of 2 · Copy copies all 9 lines
cat > leads/lead-1.txt <<'EOF'
Hi, I'm Dana. We had hail last week and I think a few shingles are gone on the back side. Can someone come take a look? Mornings are best.
EOF
cat > leads/lead-2.txt <<'EOF'
Hey there! We help roofing companies get more exclusive leads. Want a quick call this week?
EOF
cat > leads/lead-3.txt <<'EOF'Example (made up): three leads · Part 2 of 2 · Copy copies all 9 lines
You guys put our roof on in the spring and there's a drip by the chimney when it rains hard. Who do I talk to?
EOFA 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.
The pattern is claude mcp add --transport http <a name> <the app's MCP address>. Get the address from the app's own help pages. Here's the real example from Claude Code's docs, connecting Notion:
Add a remote MCP server (run in your terminal, not inside Claude)
claude mcp add --transport http notion https://mcp.notion.com/mcp--scope user to use it in every folder, or --scope project to share it with your team through a .mcp.json file.Start claude, type /mcp, pick the server and follow the steps in your browser. Adding a server doesn't check your login, so this is where you find out it works.
Prompt: a safe first test
Using the notion connection, list the names of my five most recent pages. Read only: don't create, edit or delete anything.Run logout before remove if you want the login cancelled on the app's side too.
Manage servers
claude mcp list
claude mcp get notion
claude mcp logout notion
claude mcp remove notionA 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.
Save this as .claude/skills/lead-reply/SKILL.md. The --- block has to be the very first line of the file. Change the steps to how your office really answers a lead.
.claude/skills/lead-reply/SKILL.md · Part 1 of 2 · Copy copies all 18 lines
---
name: lead-reply
description: How we reply to a new job lead (a quote, an inspection, a repair). Use whenever drafting a first reply to a new lead.
---
# Replying to a new job lead
1. Thank them by first name if they gave one. If not, just say hi.
2. Repeat back what they need in one sentence, in their words.
3. Ask one question that gets the visit booked: "Does Tuesday or Wednesday morning work for someone to come look?"
4. Sign off with the company name and the office phone line, written as <office phone>..claude/skills/lead-reply/SKILL.md · Part 2 of 2 · Copy copies all 18 lines
Never:
- quote a price or a timeline
- say "guaranteed"
- add anything they didn't tell us
Keep it under 90 words. Write it like the person at the front desk talks.Type /lead-reply to run it yourself, or just ask for a reply to a lead and Claude loads it from the description.
Prompt: try the skill
/lead-reply Draft a reply to leads/lead-1.txt. Show it to me; don't save or send anything.A skill can tell Claude to run code. Anthropic's advice is to use skills you wrote yourself or got from Anthropic, and to read every file in anybody else's before you use it.
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.
Read-only: it can look at the leads and nothing else.
.claude/agents/lead-router.md · Part 1 of 2 · Copy copies all 22 lines
---
name: lead-router
description: Reads one incoming lead or message and labels it so it goes the right way. Use for every new message in leads/ before anything else touches it.
tools: Read, Glob
model: inherit
---
You sort incoming messages for a home-service business. You never reply to anyone and never change any file.
For each message you're given, pick exactly one label:
- NEW JOB: someone wants a quote, an inspection or work done
- CUSTOMER: an existing customer with a question, a warranty issue or a complaint
- VENDOR: someone selling us something (leads, ads, software, SEO)
- SPAM: junk, or nothing we can act on.claude/agents/lead-router.md · Part 2 of 2 · Copy copies all 22 lines
Answer in this shape, one block per message:
File: <file name>
Label: <one label>
Why: <one sentence, quoting the words that decided it>
Next: <NEW JOB -> reply-writer, CUSTOMER -> flag for the office, VENDOR -> ignore, SPAM -> ignore>
If you can't tell, say UNSURE and why. Don't guess.It can write, but only drafts.
.claude/agents/reply-writer.md
---
name: reply-writer
description: Drafts a reply to one NEW JOB lead, following the lead-reply skill. Use after the lead-router labels a message NEW JOB.
tools: Read, Write, Glob
model: inherit
---
You draft replies for a home-service business. You only write drafts into the drafts/ folder. You never send anything.
Follow the lead-reply skill in .claude/skills/lead-reply/SKILL.md exactly.
Save each draft as drafts/<lead file name>.md and end your answer with the file you wrote.No Write or Edit tool, so it can't fix its own findings or anybody's draft. It only grades.
.claude/agents/reply-checker.md · Part 1 of 2 · Copy copies all 17 lines
---
name: reply-checker
description: Grades a drafted reply against the checklist and returns PASS or a list of fixes. Use after the reply-writer writes a draft, before anyone sees it.
tools: Read, Glob, Grep
model: inherit
---
You check drafted replies. You can't edit anything, on purpose: you only grade.
Read the draft and the lead it answers, then check every line of this list:
1. It answers what the person actually asked.
2. It asks for one next step (a time for the visit, or a call back), not five.
3. It promises nothing we can't back up: no prices, no dates, no "guaranteed".
4. It's under 90 words and sounds like a person at the office, not a brochure.
5. It has no made-up details (names, addresses, numbers) that aren't in the lead..claude/agents/reply-checker.md · Part 2 of 2 · Copy copies all 17 lines
Answer PASS, or FAIL with each failed line by number and the exact fix.Quit Claude Code and run claude again in the same folder. A running session doesn't notice a brand-new agents folder; after that, changes to the files show up within a few seconds.
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.
Prompt: route the leads
Use the lead-router subagent on every file in leads/ and show me its labels. Don't write or send anything.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.
Prompt: router, then writer
Use the lead-router subagent on leads/lead-1.txt. Only if it says NEW JOB, use the reply-writer subagent to draft the reply into drafts/. Show me the label and the draft. Don't send anything.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.
Prompt: three at once
In parallel using separate subagents, run the lead-router on each file in leads/, one subagent per file. Then stitch the results into one table in drafts/summary.md: file, label, reason, next step. Don't send anything.Every subagent uses your plan's usage while it runs, and each one's results come back into your main chat. Claude Code runs up to 20 at once by default. Ten small jobs at once is fine; a hundred, do in batches.
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.
Real lead messages have people's names, phone numbers and addresses in them. Decide what you're OK sending to Claude before you point it at real ones, keep the router read-only, and keep every reply a draft that a person sends.
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.
Prompt: let the boss split it
You're the boss on this one. Look at everything in leads/ and work out what each lead needs. Show me your plan first: which subagent gets which lead and in what order. After I say yes, hand the work out, then give me one summary of what got done and what needs a person. Drafts only, nothing sent.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.
Prompt: write, check, fix
Use the reply-writer subagent to draft a reply to leads/lead-1.txt. Then use the reply-checker subagent to grade the draft. If it fails, send the checker's fixes back to the reply-writer and check again. Stop after two rounds and show me the final draft and the checker's last answer. Don't send anything.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).
Checked 2026-10-31 on claude.com/pricing and the Claude Code docs. Prices change: check before you buy.
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.
claude: command not found after installingOpen a new terminal window. Still missing? The install folder isn't on your PATH yet; Claude Code's setup page has a Fix your PATH section.
Restart Claude Code. A running session doesn't see an agents folder that didn't exist when it started.
The --- block must be the first line of the file. A subagent needs a name (no colons) and a description. Unknown fields are skipped without an error, so check the spelling.
Rewrite the description to say what it does and when to use it, with the words you'd actually type. Or call it by name: /lead-reply.
Run /mcp inside Claude and log in. claude mcp add saves the server without checking your login.
Claude Code warns past 10,000 tokens and cuts off at 25,000 by default. Ask for less (one page, five records), or raise the limit with the MAX_MCP_OUTPUT_TOKENS setting.
Concurrent subagent limit reachedThe default is 20 at once. Run the job in batches.
Parallel runs and fix loops spend it. Fewer subagents at once, shorter loops (two rounds), and send each one only the files it needs.
Tighten the label definitions in lead-router.md and add an example of each. Keep the UNSURE option so it asks instead of guessing.
Its checklist is too loose. Make each line something it can check yes or no, and keep it read-only.
What each page backs up is listed at the end of the full guide.
A plug, a process, a chain, a crew, a sorter, a boss and a checker, and the checker goes in first, because nothing should grade its own homework.
Want help setting this up? Receipts Group builds these systems.
Open the full guide as a page (every file in full, plus the table of contents).