How to structure a team of AI agents like a company

Why an org chart beats a pile of agents

Most people collect agents. They end up with fourteen clever assistants, no idea which one to open, and every single one re-asking who the business is for. That is not a team, it is a drawer.

A company solves this with three boring things: a shared understanding of the business, a written process for each recurring job, and a named owner for that job. Give your agents the same three things and they stop being demos. The org chart is what turns "I have some prompts saved" into a system you can hand to somebody else.

Build it in this order

This order matters. People who start by building agents rebuild them all three weeks later when they finally write the brain.

1. 🧠 Write the brain first. One document about your business, before any agent exists. Everything downstream inherits it, and it is the difference between an agent that sounds like your company and one that sounds like a stock assistant.
2. 📋 List every recurring workflow. Not tasks, workflows: things that happen repeatedly with a recognisable start and a recognisable finish. Most solo businesses land somewhere between 25 and 60. Write them all down before grouping anything.
3. 🗂️ Group them into departments. Six is a good ceiling. If a workflow fits two departments, put it where the output lands, not where the work happens.
4. 🚦 Label each one for autonomy. Do this before you build, while you are honest. Anything that sends, spends, publishes or promises starts as human-assisted, no exceptions.
5. 📄 Write SOPs for the three that hurt most. Rank by how often it repeats multiplied by how much you dread it. Three good SOPs beat thirty stubs.
6. 🤖 Attach an agent to each SOP. Now, and only now, build the agent. The SOP is most of its instructions already, so this part is fast.
7. 🎛️ Add the conductor last. One routing agent that reads the directory and dispatches to the right worker. Useless until there are enough workers to be worth routing between.

The six departments

These six cover almost every small business or solo operation. Rename them to match how you actually talk, and take the first agent as your starting point in each.

Write the brain

The brain is one document, roughly one to three pages. Every agent reads it, so a fact you fix here is fixed everywhere. Run this prompt in any Claude chat and answer its questions properly, because this is the highest-leverage half hour in the whole build.

I am building a system of AI agents to run the repeatable parts of my business, and I need the shared knowledge base they will all read before anything else. Interview me first: ask me one question at a time about what I sell, who exactly buys it and what they are usually using instead, how I price, what my delivery process looks like, how I want to sound in writing, and the things I never want said or done in my name.

Then write the document. Structure it as: what the business is in two sentences, the offers and prices, the ideal customer and the customer I decline, the voice rules with two real examples of good and bad, the tools I actually use, the numbers that matter to me, and a hard rules section for anything that must never happen without me.

Write it in plain sentences an agent can follow literally, not marketing copy. Where my answers were vague, say so in a short open-questions list at the bottom rather than inventing something plausible. Keep the whole thing under three pages.

Then give it a home so agents can actually reach it. Pick one and stay consistent.

One SOP per workflow

The SOP is the thing that makes this a company rather than a prompt library. It is short, it is specific, and it exists whether the work is done by an agent or by a person. Six fields, every time.

Have Claude write them for you rather than typing forty SOPs by hand.

Here is a recurring workflow in my business: [describe it roughly, the messy version is fine, including the bits you always forget].

Interview me until you can write it properly. Ask about the trigger, the inputs, the exact steps in order, what the finished output looks like and where it gets saved, who should own it, and which parts genuinely need my judgement rather than my time. Ask about the failure cases too: what has gone wrong before, and what would make this unrecoverable rather than annoying.

Then write the SOP with exactly these sections: trigger, inputs, steps, definition of done, owner, autonomy level. Recommend an autonomy level of human-led, human-assisted or autonomous and defend the choice in one sentence, defaulting to human-assisted for anything that sends, spends, publishes or promises. Finish with a short list of what an agent would need access to in order to run this, so I know what to connect before I build it.

Workflow: weekly client status report
Department: clients
Trigger: every Friday 9am, or when a client asks for an update
Inputs: this week's task board export, my raw notes, last week's report
Steps:
  1. Pull wins, blockers and next steps out of the raw notes.
  2. Compare against last week's next steps and flag anything that slipped.
  3. Write it in the house format: one-line summary, then those three sections.
  4. Keep the tone plain and factual, no filler, no apologising.
  5. Save to the client folder and put the draft in front of me.
Definition of done: a drafted report in the client folder, unsent, with slipped items flagged
Owner: reporter agent
Autonomy: human-assisted (I read it before it goes out)

The three autonomy levels

Every workflow in the directory carries one of these. It is what stops a system like this from being frightening, and it is the reason you can grow it without lying awake.

• Human-led. You do the work, the agent assists inside it: research, drafting, checking, second opinions. Judgement calls, pricing, hiring, anything with taste or consequence lives here permanently.
• Human-assisted. The agent does the work and you approve the output before it leaves the building. This is where most of your system should sit, and where things stay until they have been right roughly ten times in a row.
• Autonomous. It runs and you read the log later. Earn it only for jobs that are reversible, low-stakes and boring: filing, sorting, summarising, tagging, monitoring.

The conductor agent

Once you have more than about six workers, remembering which one to open becomes its own small job. The conductor is one agent that holds the directory, works out which workflow you are describing, gathers the inputs that workflow needs, and hands off. It does no real work itself, which is exactly why it stays reliable.

Build me a reusable Claude skill called "conductor" that acts as the delegator for my agent system.

Its description should say to use it whenever I describe something I need done and have not named a specific agent. Its instructions should tell you to hold a directory of my workflows, each with its department, owner agent, autonomy level and required inputs, and to match my request to the closest workflow. If nothing matches well, say so and offer to draft a new SOP rather than improvising.

Before handing off, check I have supplied every input that workflow needs and ask me for the missing ones. Always state out loud which workflow you matched, which agent owns it, and its autonomy level, so I can catch a wrong routing before any work happens. For anything labelled human-assisted, stop at a draft and hand it to me for approval. Never merge two workflows into one run, and never do the work yourself when an owner agent exists.

Here is my directory to load: [paste your workflow list with department, owner and autonomy for each].

Build the directory you can click into

The visual directory is not decoration. It is the thing that tells you what exists, what is running itself, and where the gaps are. A database with the right columns takes about twenty minutes.

1. 🗄️ Make one database, one row per workflow. Notion, Airtable or a spreadsheet, whichever you already open daily. One row per workflow, never one row per agent, because agents come and go and workflows do not (Notion databases intro).
2. 🏷️ Add these columns. Workflow, department, trigger, owner agent, autonomy, SOP link, skill file, status, last reviewed. Nine columns, and every one of them earns its place (property types).
3. 👀 Group the board by department. That single view is the org chart. A second view filtered to autonomous shows you everything currently allowed to run without you, which is the view worth checking monthly (views and filters).
4. 📎 Attach the real files. Each row links to its SOP page and its skill folder or zip. This is what makes the directory usable by somebody who is not you, including a future assistant or contractor.
5. 🩺 Review it monthly. Fifteen minutes: what has not run, what is still human-assisted and has earned promotion, what SOP is now wrong because the business moved. A directory nobody reviews becomes fiction within a quarter.

What actually makes it run

Here is the part the impressive screen recordings skip. A directory of agents does not wake up on its own. Something has to pull the trigger, and there are only three somethings.

Proof and further reading

Agent Skills, official overview
Skill authoring best practices
Creating custom skills (Claude Help Center)
What Projects are
Create custom subagents · agents overview · memory and CLAUDE.md
Connecting Gmail, Calendar and Drive