← All guides
Business · Automation ·Claude, ChatGPT

How I Built a 35-Agent AI Marketing Workforce (Without Writing Code)

I run a marketing org where 35 AI agents handle the recurring work: reporting, briefs, insights, and more. Here is the architecture, in plain English, and how to build your first three agents this month.

I run growth marketing for a national brand, and some of the most reliable members of my extended team are not people. They are 35 AI agents I built with plain English instructions, no code, organized like a department: a chief of staff that routes the work, domain leads for analytics, content, insights, and operations, and specialists under each.

That sounds like science fiction. It is actually just one boring, powerful idea applied over and over. Here is the whole architecture, and how to start yours.

It started with one audit, not a vision

I did not sit down to build a workforce. I ran a task audit on my own week, found that reporting and first drafts were eating hours every single week, and built one saved agent for reporting. That agent worked. So I built the next one.

The workforce is what happens when you never stop doing that.

The architecture, in plain English

Three layers, same as any org:

  1. Specialists. Each agent does exactly one job: draft the weekly report, summarize competitor moves, turn raw survey data into an insight memo, write the first pass of a brief. One agent, one job, standing instructions. This is the Saved Helper idea, multiplied.
  2. Leads. Groups of specialists organized by domain: analytics, content, customer insights, operations. A lead is really just a naming system and a shared set of rules, so related agents behave consistently.
  3. A chief of staff. One coordinating setup that knows what every agent does. I describe what I need, it tells me which agent handles it, or drafts the instructions for a new one. This is the piece that makes 35 agents feel like one team instead of 35 sticky notes.

The three rules that keep it from falling apart

Every agent asks instead of guessing. The last line of every instruction set: if something is missing, ask me. An agent that invents a number is worse than no agent.

A human signs everything that leaves the building. Agents draft, people decide. Reports get sharpened by me. External anything gets human eyes. The agents give back time, they do not take over judgment.

Failures get written down. When an agent gets something wrong, the correction goes into its instructions so it cannot make the same mistake twice. That habit is the difference between a system that improves and one that decays.

What this actually changed

The honest version: recurring production work that used to consume my team’s week now shows up as drafts waiting for judgment. The team spends more time on strategy, creative, and the conversations that actually grow the business. That was the point all along. Not fewer people. More capacity for the work only people can do.

Build your first three this month

  1. Week 1: run the task audit and build one agent for your top Draft task. Reporting is the usual winner. Here is mine.
  2. Week 2: build the second for your top recurring writing task: briefs, updates, summaries.
  3. Week 3: write your first “chief of staff” note: a saved message listing what your agents do, so future-you remembers what exists.

Three agents is not 35. It does not need to be. Three agents running reliably will return more hours than most teams’ entire AI strategy, and the rest is just compounding.

Common questions

Do you need to be technical to build an agent workforce?

No. Every agent in my system is written in plain English instructions, not code. If you can write a clear job description, you can build an agent. The skill is knowing your work, not knowing software.

What does a 35-agent workforce actually mean?

Thirty-five saved, single-purpose AI setups, each with standing instructions for one job, organized under leads that route work between them. Think org chart, not robot army. Most are simple. The power is in how they connect.

Where should a normal team start?

With one agent, built for your single most repetitive task. Mine started with reporting. Prove it works for a month, then add the second. The workforce is just that habit, compounded.

How do you keep AI agents from making things up?

Every agent I run has the same standing rule: if information is missing, ask instead of guessing. That one line prevents most of the trouble. The rest is keeping a human review on anything that leaves the building.

The Newsletter

Get the free course: 5 Days to Actually Using AI.

Five short emails, one small win a day, from setup to a system. Then one practical AI idea every week. No hype, no jargon, no tech background needed.

Free forever. Unsubscribe anytime.