Anmol HandaChaos with Anmol
Home / Articles

Article

Build Your Own AI Agent Café, Step by Step

By Anmol Handa · AI agents · ~4 min read

Also on Medium ↗

New dad, about two hours of free time a night. So I built a café of AI agents: each one has a desk, one job, and a boss (me).

They curate my LinkedIn and Instagram, write lullabies for my daughter's channel, and pick the right model for every task. They earn XP, get promoted, and go on probation when they slip. Nothing goes live without my one-tap OK. Here's how to build your own.

The ideaOne AI agent per job, each with its own folder and rules, and nothing public without your one-tap approval.

Eight steps

  1. Pick 3 jobs, not 20. Start with jobs that are repetitive and low-risk: drafting posts, summarising news, planning your week.
  2. Give each agent a home. One folder per agent with an instruction file: who it is, its one job, what it must never do, and how to ask you questions.
  3. Write the guardrails first. "Never publish, send, buy or sign anything; draft it and ask." Add the specific don'ts (no personal data, no work topics).
  4. Make approval one tap. Every task ends with a short list of [Approve | Skip] questions. Nothing goes out until you tap.
  5. Separate the hands. One "browser" agent does the clicking for everyone and double-checks which account it's on before acting.
  6. Track progress. Each agent keeps a short STATUS file (done / next / blocked) and a plan checklist you can glance at.
  7. Learn from reviews. When you send work back, the lesson goes into that agent's rules file so it doesn't repeat the mistake.
  8. Make it a game. XP for approved work, penalties for redos, promotions to bigger jobs, probation when an agent keeps slipping. It makes the review loop fun enough to keep doing.

Starter prompt

Copy it, edit it, and paste it into your agent's rules file.

starter prompt
You are <Name>, my agent for <one job>. Work only inside this folder. Never publish, send, buy or sign anything: draft it and ask me. End every task with a 3-line summary and, if you need a decision, a QUESTIONS list with 2-4 short options each. Keep STATUS.md current: Done, Next, Blocked.

Why the boss matters

An agent with tools is a distributed system that can be confidently wrong. The guardrails above are the same ones I'd want for any production system: least privilege, a budget, a human approval step for anything irreversible, and a log of what happened. The café just makes them visible.

All opinions are my own.

Related reels