← Back to all free guides
A free guide by Mariah Brunner ↗ INSTAGRAM ↗ TIKTOK

How I Build
My AI Agents

Every agent running my two businesses is built the exact same way, in four parts. Here’s the full structure with real examples, plus a prompt to start your first one.

A chatbot waits for a question and gives a general answer. An agent has a job. It knows your business, it’s plugged into your real data, it follows set processes and it remembers what happened last time.

Think of each agent as an employee with one role. On my team I have a Meta media buyer, a Google Ads buyer, an email marketing strategist, a creative director and a co-founder agent. Every single one is built with the same four parts.

The 4 Parts

01

Context

Files that teach the agent your business. Not a one-line prompt. A set of documents it reads every time you work together. Every agent of mine gets four:

  • Identity: its role, how it thinks and what it never does. My media buyer’s says it optimizes for profit, not vanity numbers, and never raises a budget without a written reason.
  • Business profile: what we sell and who buys it. It also has our margins and our busy seasons.
  • Rules and targets: my cost-per-purchase targets and plain-English rules, like "cut an ad at 2x target cost with no purchases."
  • Past decisions: big calls we made before and why, so it knows the precedent.

The most common mistake is one giant prompt. If you have to re-explain your business every session, you have a chatbot, not an agent.

02

Connections

The live data the agent pulls from, so it works from what’s happening right now and not just what it was trained on. Mine connect to Shopify, Klaviyo, Meta Ads, Google Ads, my customer service platform, competitor ad libraries and a list of experts on X I want to keep up with.

In Claude, many of these are one-click connectors under Customize. For the rest, Claude Code can write the connection for you. My first agent took a weekend. Every one after that took a few hours, because the pattern repeats.

03

Workflows

The agent’s repeatable processes, written down and named, so it runs them the same way every time.

  • Daily: a morning brief. Yesterday’s numbers against the last 7 and 30 days, anything unusual, and the top 3 actions with a reason for each.
  • Weekly: a full audit of what to scale and what to cut, turned into a brief my team can act on.
  • Monthly: a strategy review of budget, what worked and what to test next.

04

Memory

The part that makes it compound. The agent saves decisions, outcomes and lessons to a file, so every session builds on the last one instead of starting from zero.

For example: "We cut this ad on May 3 because people had seen it 8 times and it cost 3x our target per purchase." When I push back on a recommendation, that pushback gets saved with the reason too, so next time it thinks differently.

Why Agents Fail

There’s a huge debate online about whether agents are overhyped. They’ve completely changed my business. But they’re not for everyone. If you’re brand new to AI, start with regular chats and Projects. And if your work fits inside Claude Cowork, stay there.

If you have a real use case and agents still aren’t working, it’s almost always one of two mistakes:

01

Building too many agents

Start with one agent doing one specific job. Add the next only when the first one is saving you real time.

02

Giving every agent every tool

Your media buyer doesn’t need your inbox, and your email agent doesn’t need your ad account. Give each agent only the tools and data its job needs. In Claude, you can set each connector tool to Always allow, Needs approval or Blocked.

What A Good Team Looks Like

Clear jobs, great memory and only the tools each agent needs. In my setup, one manager agent oversees the agents beneath it. And a human, me, always makes the final call.

Start Your First Agent

Pick one job you do over and over. Open Claude Code in an empty folder and paste this. It interviews you, then builds the four parts as files.

Build My First Agent
I want to build my first AI agent in this folder. Its job: [the task, e.g. "check my Shopify sales every morning and tell me what changed"]. Interview me first. Ask one question at a time about my business, my goals, the numbers I care about and how I make decisions for this job. Stop when you have enough. Then create these files: 1. context/identity.md: the agent's role, how it thinks and what it must never do 2. context/business.md: my business, customers and goals 3. context/rules.md: my targets and decision rules in plain English 4. workflows/daily.md: the exact steps for the daily check and what the report looks like 5. memory/decisions.md: an empty log where you'll save decisions, results and lessons 6. CLAUDE.md: tells you to read the context files at the start of every session and add to memory/decisions.md at the end Then tell me which of my tools you'd need to connect and how to connect each one. Don't change anything in my accounts. Only read and recommend.
  1. Run it by hand first. Each day, ask it to run its daily workflow. Check its work for a week or two until you trust it.
  2. Then put it on a schedule. Once it’s saving you real time, let it run on its own.
My Setup

I build every agent in Claude Code. That’s where the context files, connections, workflows and memory live. You can run an agent completely by hand from there.

OpenClaw is my optional second layer. It’s a free, open-source assistant that runs on your own computer and lets me message my agents in Telegram. They run on a schedule and send me their reports, so my media buyer’s brief is waiting every morning without me asking. To use Claude inside OpenClaw you need a Claude API key, because Anthropic doesn’t let Pro or Max logins power outside tools.

Prefer to stay inside Claude? Paid plans can now run scheduled tasks in the Claude app and in Claude Code, no extra tools needed.

You don’t need both on day one. Don’t automate something you don’t trust yet.

Want to see two of my agents up close? Here’s my Meta Media Buyer Agent and my Google Ads Agent.