The Company That Runs While You Sleep
How a small company can be built from AI agents that work together on their own, and what the humans still need to own.
Imagine going to bed with a piece of work half-planned, and waking up to find it designed, built, reviewed and tested. One short report is waiting. It asks for two decisions. Everything else is done.
That is no longer a thought experiment. Small teams are starting to work this way, and running experiments like it in a personal lab has taught one lesson above all others: the result has less to do with how clever the models are, and much more to do with how the work is organised.
Most of us still use AI as one very capable assistant. You ask, it answers, you check, you ask again. That is useful, but it isn’t a company. A company is a set of roles, rules and handoffs that keep working when nobody is watching. It is now possible to build that out of agents. Here are the patterns that seem to matter most.
Agents do the work
Humans decide at a few gates
Everything in between is written to one shared, append-only log that any person can read later.
Hire roles, not prompts
A company doesn’t have “an employee”. It has a product lead, an architect, a builder, a reviewer, a tester, and people who think about security and risk.
So give each agent a job description, not just an instruction. What does it own? What standard does it hold? And just as important: what is it never allowed to do? A builder writes code but never merges it. A reviewer judges but never writes. A tester looks for problems and is told plainly that a polite approval helps no one.1
Clear boundaries are what make it possible to stop watching.
Separate the judge from the builder
The most important structural choice is simple: never let the same agent grade its own homework.
Work runs best as a pair. One agent judges and one builds. They talk through a shared, append-only log, so every request, answer and decision is written down. The judge sets the bar and checks the work. The builder does the work and asks when something is unclear. Neither can quietly change the rules.
This mirrors something software teams learned long ago: code review works because a second pair of eyes didn’t write the code. When a decision needs expert input, a fresh agent with a narrow question and a short reading list, asked to look for flaws, tends to review a plan far more honestly than the one that wrote it.
An example. A builder agent finishes a new sign-up flow and marks it ready. The judge reads the brief again, tries the flow with an expired email link, and finds it fails silently. It sends the work back with one sentence explaining why. Nobody had to be awake for that conversation.
Write the operating model down
Human teams run on habit and hallway conversations. Agent teams have neither. Everything they know about how your company works has to be written: the stages work moves through, the gates between them, what counts as a serious defect, when to stop and when to escalate.
This turns out to be a gift. Agents follow a written process very consistently. They don’t get tired or skip a step late on a Friday. If the process is good, they will run it faithfully at three in the morning.
If the process is vague, they will be faithfully vague.
Put people at a few decision points, not every step
A common first experience is becoming the busiest person in your own company: copying messages between agents, approving every small step, answering questions only you can answer. The work is automated, and you’ve become the courier.
A calmer shape is to appear at a handful of points: the direction, the design, the architecture, the test report and the release. Between those points, agents have clear authority to keep going, and they only reach out for a real blocker or a serious finding.
The goal isn’t to remove the person. It is to spend human attention only where it changes the outcome.
Check the value, not just the safety
One failure is easy to miss. An agent team builds a feature. Every test is green. Nothing is broken. And the feature reaches no one, because the data it was waiting for never arrives.
Safe and working are different things. Many automated checks prove the first and say little about the second. A useful habit: if a test would also pass against the old code, it isn’t evidence. Run the real thing against real data and confirm something useful actually happened.
Give the company budgets and brakes
Autonomy without limits is just a faster way to make expensive mistakes. Each agent team needs a budget, a cap on retries and a rule to pause when it’s idle or waiting on a person. Anything that touches money, customers or the outside world should wait for a human to say yes.
These brakes are what make it safe to take your hands off the wheel.
What stays with people
Some things stay with the founder, however capable the agents become.
- Taste. Deciding what is worth building, and what good looks like.
- Irreversible decisions. Anything that can’t easily be undone.
- Accountability. Customers and regulators deal with people and companies, not with agents.
- Talking to customers. Understanding what a real person struggles with is still human work.
The new shape of a company
When building gets cheaper, a viable company can get much smaller. A few people with a well-run team of agents can now take on work that once needed a much larger group.
The companies that do this well probably won’t be the ones with the most agents. They’ll be the ones with the clearest intent, the best-written operating model and the patience to check what their agents actually achieved.
The agents can keep the company running overnight. People still decide what the company is for.