The company that runs itself
One afternoon this week I built a company.
Not a real one — an AI one. A CEO agent, a content strategist, a writer, an engineer. Roles, reporting lines, budgets. I assigned tasks. Then I watched.
Six tasks ran end-to-end. Some needed a nudge — a comment to unstick a blocked agent. One needed a decision only I could make. The rest just moved through: assigned, picked up, worked, completed, reported.
The question I kept asking myself was not “is this impressive?” — though it was. It was: what would I actually hand off, and what would I keep?
That question used to be easier. The constraint was capability — could the thing do what you needed? If yes, delegate. If no, keep it yourself or hire someone who could.
AI agents have dissolved that constraint faster than most people have updated their mental model. The capability question is largely answered. Agents can write, research, communicate, execute code, manage tasks, brief other agents, and report back. The gap between “what AI can do” and “what my business actually needs done” is closing week by week.
The new question is harder: what do you want to be in the loop for?
Most operators don’t have a clean answer, because they’ve never had to have one. The default was participation everywhere — not because they chose it, but because the tools required it. You were in every email thread because you were the one sending the emails. You were in every decision because you were the one executing the decision. The loop wasn’t a choice; it was the only available workflow.
Now the loop is optional.
The tool I used — Paperclip — has an interesting design choice embedded in it: agents can do almost anything autonomously, except hire new agents. That one action always requires a human to approve. Someone has to say: yes, bring this entity into the organisation.
I thought that was a quirk at first. The more I’ve sat with it, the more I think it’s the right line to draw. Everything downstream of hiring — task assignment, execution, reporting, inter-agent coordination — can be automated. The decision to expand the organisation’s capacity is different. It’s irreversible in a way that most task decisions aren’t. It shapes what the system becomes.
The designers had to pick a line somewhere. They picked it at hiring. That choice says something.
Every business has a version of this question: where does the human decision live?
For most of business history, the answer was “everywhere” — because the cost of not being in the loop was too high. You couldn’t trust a system to handle it without you because there wasn’t a system, just a set of people and processes that needed management.
The bottleneck in any operation is not capability. It never was. It’s attention — the finite resource that owners and operators apply to their businesses. What you pay attention to is the actual strategic decision. Everything else is execution.
AI agents don’t change what’s worth paying attention to. They change how much of the rest you have to touch.
The company that runs itself isn’t a destination. It’s a direction. Every task you move off your plate and onto an agent’s is a unit of attention freed for something only you can do.
The interesting question isn’t whether that’s possible. It is.
The interesting question is: what are you saving your attention for?