Agents, in plain English
⏱ 15 min4 steps
This one runs in , in . If you'd rather stay in the , see One agent, one job.
A normal chat is you asking and Claude answering. An is Claude given a goal and some tools. It plans, does a step, checks the result and keeps going until the job is done.
Think of a busy family kitchen at Thanksgiving. One person on the turkey, one on sides, one setting the table, and someone tasting everything before it goes out.
Below, one agent checks your plan and a second one argues with it. If Claude skips the second helper, that's normal.
Claude Code Claude, working inside Terminal. It can make files and run things on your Mac, with a safety checker watching what it does. All wordsTerminal An app on your Mac where you type instructions instead of clicking on things. Claude Code works inside it. All wordsClaude app Claude on your phone or computer, where you chat with it by typing. claude.ai is the same thing in a web browser. All wordsagent Claude given a goal and some tools. It plans, does a step, checks the result and keeps going until the job is done. All wordsBefore this, you'll want:
Steps
Go back to your first-job conversation
-
Open Terminal (Cmd+SpaceWindows key, type Terminal, press Enter).
-
Go to your first-job . Type this and press Enter:
cd ~/Backyard/first-jobcd ~\Backyard\first-job -
Check the says first-job.
-
Pick up your last conversation in this folder. Type this and press Enter:
claude -c
You should see:
Claude Code opens, with your earlier conversation above the line where you type.
That's not what I see
- It says "cd: no such file or directoryCannot find path … because it does not exist".
- The first-job folder isn't there. It comes from Your first real conversation with Claude, so do that piece first, then come back.
- It says "zsh: command not found: claudeThe term 'claude' is not recognized".
- This Terminal window doesn't know about Claude Code. Close the window with Cmd+WCtrl+Shift+W, open a new one and try again. If it still says that, go back to Claude Code and .install Put a new program on your computer so you can use it. All wordssign in Tell Claude Code which Claude account is yours, so it knows you've paid for it. All words
-
Watch Claude work through a job on its own
The tools an agent gets are things like reading files, browsing the web or sending a message. It works through the steps by itself. This asks Claude to check your plan that way, one step at a time, without changing any files.
- Copy the prompt below.
Read plan.txt. Work through this as an agent: first tell me your own plan for checking it, then check each step of my plan against what I told you about my life, one at a time, and keep going until you're done. Show me each step as you go. Don't change any files. At the end, give me the three biggest problems you found.
- Paste it into Claude Code with Cmd+VCtrl+V and press Enter.
- Watch what happens.
Claude makes a plan, does a step, looks at what it found, and decides what to do next, without you asking each time. That loop is what makes it an agent. Notice the prompt also says "don't change any files." Starting with an agent that only reads and reports is the safest way in. In , the safety checker treats a line like that as a boundary, and blocks changes that would break it.
prompt What you type to Claude. A good one says what you want, how to work and what never to do. All wordsauto mode Claude works without asking, and a separate safety checker blocks anything risky. All wordsYou should see:
Claude reads plan.txt and tells you its plan for checking it. Then it goes through your steps one at a time, and ends with the three biggest problems it found.
That's not what I see
- Claude says it can't find plan.txt.
- Claude Code is in a different folder. Type , then ~/Backyard/first-job~\Backyard\first-job, check the prompt line, and type claude -c to pick up where you left off./exit Typed in Claude Code, it ends the conversation and takes you back to Terminal. All wordscd The command that moves Terminal into a folder. Type cd, a space and the folder's name. It's short for change directory, an old word for folder. All words
- Claude asked to change a file, or changed one.
- If it's asking, choose No. Either way, type: Please don't change anything, only read and report. Letting it look before it touches is the first rule of working with agents.
Get a second opinion that tries to prove it wrong
Several agents means several working at once, each on its own piece. This prompt asks Claude Code to start a second helper, with its own fresh look at the plan, whose only job is to prove the first answer wrong.
- Copy the prompt below.
Now start a second helper whose only job is to prove your three problems wrong. It should read plan.txt fresh, without seeing your reasons first. Then show me, side by side, where the two of you agree and where you disagree. Don't change any files.
- Paste it into Claude Code with Cmd+VCtrl+V and press Enter.
A second pass whose job is to find mistakes catches most confident nonsense. It's the person tasting everything before it leaves the kitchen.
You should see:
A line with Agent while the second helper works. When it's done, Claude shows where the two of them agree and where they disagree, side by side.
That's not what I see
- Claude answered by itself instead of starting a second helper.
- That's fine. The point is a second look that argues back. If it agreed with everything, ask it: Argue the other side. What would someone who thinks this plan is fine say?
- The two answers disagree and I don't know which is right.
- That's the useful part. You know your own life better than either of them. Pick the one that matches what you know, and tell Claude why.
Write one job with a clear finish line
When you're ready to hand off work, these are the rules I follow:
- One job per agent, with a clear . "Check every page for broken links and list them" works. "Improve the website" doesn't.
- Let it look before it touches. Start with agents that read and report, before any that change things.
- You approve anything that goes out. Emails, posts, payments and published pages wait for your yes.
- Make it show its work. Ask for sources, so you check claims instead of trusting them.
- Have something check the work. That's what you did in step 3.
- Only automate what you've done by hand. Until you've done a task with Claude a few times, you don't know what "good" looks like.
- Write one sentence for a job in your own week that has a clear finish line.
- Keep it somewhere you'll see it.
You're not handing it off today. You're learning to describe it.
finish line How you'll know a job is done. "List every broken link" has one. "Make it better" doesn't. All wordsThat's not what I see
- My job sounds too big.
- If you can't tell when it's done, it's too big. "Improve my website" has no end. "Check every page of my website for links that don't work and list them" does.
Why you don't need them yet
- One good conversation with Claude handles most everyday work.
- Agents add cost, moving parts and more ways to go wrong.
- If one Claude can't give you a good result, ten won't either.
- Learning to describe a job and check the result is the core skill. Agents depend on it.
Why you will want them
- Work that repeats: the same report every Monday, the same inbox sort every morning.
- Work that splits up: reviewing five parts of a website at once.
- Work that should happen while you sleep.
- A second opinion: one agent does the work, another tries to prove it wrong.
My own projects run on agents now. Every day they send me a report on everything I'm working on, sort an inbox, and check every change before it goes live. I did each of those jobs by hand first, many times, before I handed it off.
Your challenge
Write down three jobs you do every week. Next to each, write how many times you've done it with Claude. Only the ones you've done three times or more are ready to hand to an agent. Keep the list; you'll use it later.
Last checked 2026-10-17
Connect and automate › 3. Agents
Agents, in plain English
One-page card · print it or keep it open
What you'll do
This one runs in Claude Code, in Terminal. If you'd rather stay in the Claude app, see One agent, one job.
The steps
- Go back to your first-job conversation
- Watch Claude work through a job on its own
Read plan.txt. Work through this as an agent: first tell me your own plan for checking it, then check each step of my plan against what I told you about my life, one at a time, and keep going until you're done. Show me each step as you go. Don't change any files. At the end, give me the three biggest problems you found.
- Get a second opinion that tries to prove it wrong
Now start a second helper whose only job is to prove your three problems wrong. It should read plan.txt fresh, without seeing your reasons first. Then show me, side by side, where the two of you agree and where you disagree. Don't change any files.
- Write one job with a clear finish line
If it goes wrong
- It says "cd: no such file or directoryCannot find path … because it does not exist". The first-job folder isn't there. It comes from Your first real conversation with Claude, so do that piece first, then come back.
- Claude says it can't find plan.txt. Claude Code is in a different folder. Type /exit, then cd ~/Backyard/first-job~\Backyard\first-job, check the prompt line, and type claude -c to pick up where you left off.
- Claude answered by itself instead of starting a second helper. That's fine. The point is a second look that argues back. If it agreed with everything, ask it: Argue the other side. What would someone who thinks this plan is fine say?
- My job sounds too big. If you can't tell when it's done, it's too big. "Improve my website" has no end. "Check every page of my website for links that don't work and list them" does.
Words
- agent: Claude given a goal and some tools. It plans, does a step, checks the result and keeps going until the job is done.
- finish line: How you'll know a job is done. "List every broken link" has one. "Make it better" doesn't.