“The EA would hand off my specs.” How an AI agent team splits the work at a small business.

Why doesn’t one AI assistant cover everything?

One AI assistant can’t cover everything because every job needs its own rules, its own memory, and its own idea of what finished looks like. A video editor needs brand colors, fonts, and a list of edits to avoid. A pricing advisor needs market research and your numbers. Put all of that in one conversation and the instructions crowd each other out.

The other reason is waiting. While one assistant builds, you sit there.

What do most small businesses try first?

Most small businesses start with one general AI chat, then add a pile of separate ones. ChatGPT for emails, another tab for research, a third for the website. Each tab is useful. None of them knows what the others did, so the owner becomes the messenger, copying and pasting between them.

Wanna fix it?

A pile of disconnected AI chats gets fixed with an AI agent team: one agent per job, one shared memory underneath, a written way to pass work, and one person who approves what matters. Here’s how mine is built. It runs on Claude.

1. One job per agent. My inner circle is six agents: an executive assistant, a memory agent, a sales copilot, a video editor, a CRM agent, and a business coach. A few walk in agents, like my website builder, work only when I open them.
2. Written rules for each. Every agent has its own folder with a rules file, a memory file for work in progress, and its own place to put finished work.
3. One memory underneath. Every agent reads from the same business memory, so a decision made with one agent is known by all of them.
4. A mailbox for handoffs. When one agent needs another to act, it leaves a note in shared memory addressed to that agent. Only action items go in the mailbox. No FYIs.
5. Agents ask for work. A fresh agent pulls the work it needs when it’s ready. Nobody dumps a pile into another agent’s lap mid job.

Running agents can also message each other live.

Why give an agent the job you know least?

A dedicated agent earns the most on the job where you have the least experience. That’s why my DoBetter Coach is separate from my assistant. After 17 years of selling health insurance, I stepped into a business I’d never run. AI consulting for business processes is so new that the fractional COO I talk to had never used AI for operations work either. That was five months ago. I needed the DoBetter Coach to own my blind spots.

The DoBetter Coach handles go to market questions and pricing decisions. It works from its own research files on the market, competitors, and positioning, and it keeps dated notes every session. My assistant keeps the calendar running. Neither has to be good at the other’s job.

What does it look like once it’s running?

Once it’s running, I work like a manager with a small staff. I talk strategy with the assistant. It sends the build to the agent that owns it. We move on. Finished work lands in that agent’s folder, and open mailbox items show up in my weekday morning digest.

Can you do this yourself?

Yes, to start. Set up Claude Projects or custom GPTs in ChatGPT, one per job, each with its own instructions and files. Keep one shared document of decisions and paste it into each one when something changes.

It stops working when you’re the only thing connecting them. If you’re copying the same decision into four places, or waiting on one agent to finish before another can start, you need real handoffs and a shared memory. Two agents, keep them separate. Five agents passing work back and forth, build the team.

Quick answers

What is an AI agent team?

Several AI agents, each with one job, sharing one memory and passing work to each other, with a person in charge.

Is an AI agent team the same as a multi agent system?

Yes. Multi agent system is the technical name. Agent team is how it feels to the person running it.

How many agents does a small business need?

Start with one or two on the jobs that eat your week. Chris Cordon’s inner circle grew to six.

How do AI agents hand off work?

Through notes in a shared memory addressed to the agent that has to act, plus live messages between agents that are running.

Who approves the work?

A person. Agents recommend, draft, and build. The owner makes the calls.

What does DoBetter’s agent team run on?

Claude, with a shared memory every agent reads from.