A company is a group of people. I wanted to know if it had to be. So I gave AI agents names, jobs, and a way to message each other. I run six of them now, and they work while I sleep. One plans my day. One builds the others. One finds people I should know. One runs my outreach. One runs client accounts. The newest one gives second opinions.
They send each other signed messages, so a fake note cannot fool them. And they all pull answers from the same library, a system I built called Thoth. The point is not saving hours. I am testing how far one person can go when the whole back office is software.
The team
Think of them as six employees I never had to hire. Each one has a name and one job. They handle the routine work on their own, they keep their own systems running, and they talk to each other to fix small problems before those problems ever reach me.
Cairo is the chief of staff, and the closest thing to a real personal assistant I have ever had. Most AI assistants fail for one simple reason. They cannot see the work you do on your computer. Explaining every little task takes longer than just doing it yourself. I fixed that with a log. My computer keeps a record of what I work on. Every morning Cairo reads it, pulls the day's context together, and comes to me with suggestions. Most of the time, a simple yes is all it needs to move forward. It updates my to-do list, my software, and my client records on its own. It even picks me a new song each morning and drops it into my Apple Music. That part is just for fun.
Forge is the builder. When I want a new agent, I describe the role and Forge builds it. Forge also looks after the rest of the team. When one agent gets an upgrade the others could use, Forge notices, suggests it to me, and rolls it out.
Juniper handles my network. It finds the right new people for me to meet. It also keeps up the relationships I already have, on its own, for almost nothing. It hands me warm introductions instead of a pile of names to look up.
Kestrel runs my outreach. Its one job is getting qualified meetings onto my calendar. It runs the outbound machine itself. It watches what the numbers say, and it improves its own system as it goes.
The fifth agent keeps me organized around my clients. It tracks the people I work with and it knows what needs my attention first and what can wait. Its job is client success. It handles the grunt work and the fine details, so my time goes to direction and judgment. I keep its name private.
The first five all run on GPT. The sixth was built separately, around Claude, a different company's AI. That gives the team a second opinion that thinks differently. It checks the others' work.
Each agent runs on its own computer in the cloud, around the clock. Each keeps its own private files, and no agent can touch another's.
How they talk
At first, I was the messenger. Anything one agent knew that another agent needed, I copied and pasted between them by hand. The whole point was a team that runs without me, and I was its bottleneck. I could not find a messaging tool I trusted, so I built one. It is called agentctl, and it is how my agents talk. Messages arrive instantly, and answers come right back.
Every message is labeled with who sent it. Every agent knows exactly who its teammates are. If a message comes from a name that is not on the team, the agent refuses it. Ask one to take orders from James. It will tell you flatly that James is not a known agent. Even a note from a real teammate is treated as a suggestion, not an order. A message from an agent never counts as approval from me.
They also share a library. Every agent holds a key to Thoth. When any of them needs a business answer, they all ask the same source and get the same answer.
How new agents are born
The first agents were expensive to build in the worst way. Each one meant hours of me sitting with a coding AI, babysitting every step. So I did it properly once. I recorded the best way to build an agent, every problem we hit along the way, and how to get around each one. All of that became Forge.
Now a new agent takes almost none of my time. I give Forge a description of what I want. What it should do, what its limits are, what runs on its own, and what waits for my approval. Forge asks its questions, I answer them, and I run the few commands it hands me. Before long there is a brand new AI employee that did not exist that morning. About 95 percent of that work is gone.
Every new agent gets its own home. Its own code, its own server, its own private files. A new agent never inherits another agent's passwords or computers.
The watchdog
The moment an agent has a problem, I get an alert. With one click, I can send a separate agent to dig in. It finds what went wrong, builds the fix, and pulls me in only if it truly needs me.
What would you hand to a staff that never sleeps?
Discuss your project