I stopped asking 1 chatbot everything. I gave each Ai agent 1 job.

01The change was in my head, not in the tools

Think about a real company. You do not hire 1 person to do everything. You hire someone for marketing, someone to book the calls, someone to close them, someone to read the numbers. Then someone who ties them together and decides what matters this week.

I built that. 1 agent takes what I want and hands the work out. 1 watches the market. 1 drafts the outreach that would otherwise sit on my list until Friday. 1 handles everything client facing. 1 grades the work of the others and flags what needs my eyes. They all read my email, my messages, my pipeline, my costs.

Stop asking what the tool can do. Ask what job you are hiring for.

That 1 question is the whole shift. Ask what job you are hiring for, and what that person would need on day 1, and the build stops being a mystery. You already know how to write a job. You have done it for people.

02The 6 roles I actually run

"A team of agents" means nothing until somebody lists the roles. Mine is 6.

Number 6 is the least obvious and the most useful, and it took me the longest to build. More on it below.

Not 1 of these works alone. The tools underneath them do not either, which is why I wrote the 5 Ai tools I pay for and why none of them work alone. A team is the point. A single clever tool is not.

03The 3 things every agent needs

Every agent that worked for me had all 3 of these. Every agent that embarrassed me was missing 1, and afterwards I could always tell which one.

The third one is the part people skip. An agent with no line into your real systems is a polished consultant who has never seen your business. My agents read from 1 company brain that holds our facts, and the assistants sit inside Telegram, where I already type all day.

A brain with no hands can only give advice.

04What actually happened, honestly

My first try was too ambitious. I stood up the whole team at once, because the whole team makes the good screenshot. I got 5 agents that each half worked, and fixing them together was miserable. I could never tell if a bad answer came from the agent in front of me or from the mess it was handed.

So I started again the other way. 1 agent, fed 10 real cases, tightened until it was solid. Then a second one wired to the first, tested at the join on its own. 1 good link at a time. It felt slower and it was far faster.

Here is the most useful thing I learned, and nobody talks about it. An agent has to hand the next agent clean, tidy information. Not a vague yes. Not a paragraph the next one has to work out. Sloppy handoffs are the number 1 reason these teams fall over, and they fail quietly, which is worse.

So when the end of the chain gives you something bad, do not blame the last agent. 9 times out of 10 it did a fair job with the rubbish it was handed 2 steps before.

Before any of this touched a real person, I threw awful inputs at it on purpose. A 1 word message. A message in another language. An angry complaint. A lead that was nothing to do with us. I was not checking for clever answers. I was checking whether it stopped and called a human instead of making something up.

A good agent is not one that never gets confused. It is one that knows it is confused and calls a person.

05The agent that grades the other agents

This is the piece that took the whole thing from interesting to something I trust, and almost nobody builds it, because it is not the fun part.

1 agent whose only job is to mark the work of the others against a written standard, and send weak work back before a human sees it. Not a vague quality check. A real list of what good looks like for each type of work, written by me.

The work gets better instead of slipping. Without a grader you get a pile of okay work and you become the quality control department. With it, weak work is caught and redone overnight, and what reaches me has already survived a critic.

Build this one second, right after your first agent works. Not last. It is the difference between a system that makes more work for you and one that makes less.

06Start with 1 agent, then hire the second

First, what it did not do. It did not replace anyone. It took away the production layer: the drafting, the sorting, the routing, the first pass. It never touched judgment, taste, relationships and the decisions that carry real consequences.

It also did not run itself from day 1. The early weeks were honest work. I spent more time correcting it than it saved me. That is normal, and the owners who quit at that point quit right before it pays them back.

And the gate. The agents work through the night, line up whatever needs a decision, and I tap yes in the morning. Anything that sends money, sends a message to a real person, or cannot be undone waits for that tap. Everything cheap and reversible runs free. That 1 choice makes an agent team a help, not a risk.

My company is a marketing company. In a normal service firm the same shape shows up with different labels. An intake agent that catches and sorts every enquiry the minute it lands. A drafting agent that builds the documents for a person to check. A follow up agent that chases what everyone hates chasing. And a boss agent that puts the few real decisions in front of the owner. In almost every service business the owner is the bottleneck, so build the roles you would hire first.

Most people will spend this year asking 1 chatbot 1 question at a time. That beats nothing. The people pulling ahead are building a team. 1 owner with an org chart of agents, shipping like a company several times the size. You do not need permission or a technical background. Name the roles, hand each one its facts, put a gate on anything that cannot be undone, and start with agent 1.

07Questions, answered