How to run a one person company with AI agents
The setup I use to run my own company with AI agents: what each agent owns, how parallel sessions work, and where a human still has to decide.
Short answer
- Agents do not replace you. They replace the part of your day that is typing, formatting, searching and repeating.
- Give every agent one clear job, one place to write, and one set of files it is allowed to touch.
- Run a few sessions side by side instead of one long chat, and keep the final decision with yourself.
Start from the work, not from the tools
Most people start by picking a tool and then look for something to automate. That is backwards. Write down what you actually did last week, hour by hour. You will see the same five or six blocks over and over: research, drafting, follow ups, reporting, admin, and building.
Each of those blocks is a candidate for an agent. Not because the agent is smart, but because the block is repeatable and the output has a shape you can describe.
One agent, one job
A single agent that does everything gives you vague results. Split the work so that each agent has a narrow job, a defined output, and a place where that output lands.
- A research agent that reads and returns notes in a fixed format.
- A writing agent that only drafts, never publishes.
- A build agent that works in your codebase with rules about what it may change.
- An ops agent that handles recurring admin, like turning notes into tasks.
Run sessions in parallel
The real speed comes from parallel work. While one session is building, another is researching, and a third is drafting the follow up email. You become the person who reviews and merges, not the person who waits.
This only works if the sessions do not fight over the same files. Keep clear boundaries, and check in changes often so you can throw away a bad branch without losing anything.
Where you still have to decide
Agents are good at producing options. They are bad at knowing which option fits your business, your clients, and the promise you already made in public. Pricing, positioning, scope, and anything a client will read with their name on it stays with you.
That is also the honest limit of this setup. It gives one person the output of a small studio. It does not give one person a second opinion that actually cares.
Common questions
Do I need to know how to code to run agents like this?
No for research, writing and ops work. Yes, at a basic level, if you want agents to build and change software for you. You do not need to write the code, but you need to read it well enough to know when something looks wrong.
How many agents is too many?
If you cannot say in one sentence what an agent owns, you have too many. Most solo operators run well with three or four.