Hire your first two AI employees in Claude Code in one evening

The problem

It's 10pm and you're explaining your prices to Claude for the third time this week. The payment reminder it wrote for you offered a discount you don't give. You watched a video about building an "AI team" in Claude Code and closed it thinking you'd need somebody's scripts to pull it off.

Steps

Quick map before you start, because every piece of an "AI team" already has a name inside Claude Code. Your company file is a CLAUDE.md. Each employee is a subagent: one small file with a job description. An employee's notebook is the memory setting on that file. The fences are permissions. The clock is a scheduled task. All of it is built in, y ya.

1. Make the office. Create a folder for your business and put one file in it with your offers, prices and the rules you never break. Open Claude Code in that folder, either in the Code tab of the desktop app or in your terminal.
2. Write the company file and hire employee #1. Paste the prompt below. Claude interviews you one question at a time, then shows you two files before saving anything: your CLAUDE.md (who you are, your rules, how you sound) and your first employee's file in .claude/agents/. Read both. Say yes when they sound like your business.
3. Check that it loaded. Type /context and look for your CLAUDE.md under Memory files. Type /memory any time you want to open and edit it yourself. If this was the very first agent file in a brand-new agents folder, close Claude Code and open it again once; after that, new and edited agent files get picked up on their own within a few seconds.
4. Give employee #1 a real job. Ask for something it should own, and call it by name so you know it was the one who did it, like: Use the follow-up-writer to draft a reminder for a client whose second payment is late. Claude reads each employee's description to decide who gets which job, so if the wrong one picks it up, the fix is almost always a clearer one-line description.
5. Hire employee #2 and set up the handoff. Ask Claude in plain words: "Create a project subagent called bookkeeper. Its job: read my payments file and tell me who owes what. Tools: Read only. Give it memory: project. It never writes to clients." Then give one sentence that uses both: "Have the bookkeeper check who owes on the payment plan, then have the follow-up-writer draft a reminder for each one." That's a handoff. Each employee works in its own separate conversation and only the result comes back to you, so your main chat doesn't fill up with the whole payments file.
6. Put up the fences. There are two kinds, and you want both.
   • Per employee: the tools: line in each agent file is a real list. If it's not on the list, that employee can't use it. Two catches. If you give an employee Bash (the command line), it can still write files through commands. And memory: turns on Read, Write and Edit automatically so the employee can keep its notebook, so a "read-only" employee with a notebook can write files.
   • For the whole office: type /permissions, or ask Claude to add these rules to .claude/settings.json in your business folder. Rules are checked deny first, then ask, then allow, and a deny always wins.

   {
     "permissions": {
       "ask": ["mcp__*", "WebFetch"],
       "deny": ["Bash(rm *)", "PowerShell(Remove-Item *)", "Read(./.env)"]
     }
   }

   In plain words: ask me before using any connected app (that's where sending and spending live) or fetching a web page, and never delete files or read my secrets file. One honest caveat from the docs: a command rule like Bash(rm *) catches the way Claude usually writes the command, and it isn't a lock on every possible way to delete. Writing "never delete" in your CLAUDE.md is a request. The permission rule is what Claude Code actually enforces, and a hook is the next level up when you need a guard that never forgets.

7. Give one employee a clock. In the desktop app, open the Code tab, click Routines in the sidebar (it can also be under the sidebar's More menu), click New routine and choose Local. Fill in Name, Description, Instructions and Schedule, and pick your business folder. For instructions, write the same sentence from step 5 plus "Save the drafts in a drafts folder. Do not send anything." Pick Weekly and a day. Then click Run now once, watch for permission prompts, and choose "always allow" for the ones that are safe so the Monday run doesn't stall waiting for you. It only runs while the app is open and your computer is awake; if it was asleep, it catches up once with the most recent missed run.
8. Hire the next one by asking. You don't need a special menu. In recent versions, /agents just reminds you to ask Claude or edit the .claude/agents/ folder, so hiring is a sentence: "Create a project subagent called [ROLE] whose job is [ONE LINE]. Tools: [ONLY WHAT IT NEEDS]." If you hire a lot, you can even make a recruiter: an employee whose only job is writing new agent files the way you like them.
9. Later, not tonight. Three things you'll hear about, with their fine print:
   • A shared task board exists only in agent teams, which are experimental and off by default, turned on with the setting CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS. It's a text task list, not a visual board, and the docs say teams use significantly more tokens than a single session.
   • Cloud routines (/schedule) keep running with your laptop closed, but they're a research preview, run no more often than once an hour, work from a fresh copy of a GitHub repo instead of your local files, and don't stop to ask permission. Connectors you leave on can write without asking.
   • Remote Control (/remote-control) lets you check on a session from the Claude app on your phone, as long as your computer stays on and Claude Code keeps running.

Every employee also uses your plan's usage. Each one runs its own conversation and those requests count toward the same limits as yours, so hire for a real job, not to fill an org chart.