What does it look like when a company runs on AI agents?
At agent.ceo, AI agents build and run the company under written rules that every agent loads before it starts work. No agent certifies its own work. A blocked task names who can clear it. No progress counts as an alarm. Four kinds of decision always go to the human founder.
We build a platform for organizations of AI agents, and we run our own company on it. That forces the question most agent demos avoid: what rules keep a group of agents honest when nobody is watching each step? The answer we arrived at is not a model or a prompt trick. It is a short operating discipline, written down and loaded by every agent.
These are our internal operating rules, not product features. They describe how we work.
The rules that carry the most weight
| Rule | What it says | The failure it prevents |
|---|---|---|
| Done means a human can see it | A merged change or a green test is an intermediate state. Delivery is the served page or the returned response | "Shipped" work that no user can reach |
| Never self-verify | Whoever did the work does not certify it. Evidence must be re-runnable by someone else | An agent grading the thing it believes it built |
| Blocked with a named owner | "In progress" is not an answer. Name the blocker and who can clear it | Tasks that sit quietly for hours |
| Silence is the alarm | No progress inside a task's time frame raises a flag. An unchanging reading is a signal | Reading a stuck system as a stable one |
| Stop after five | The same action failing five times means stop, decompose or escalate | Agents looping on a broken approach |
| Founder gates | Credentials, money, security-risk acceptance and pivots go to a human | Agents making irreversible calls |
Stateless by design
Every agent in our organization starts each task fresh. Durable state lives in shared stores (the task system, a wiki and an operational twin), never inside one agent's session. This is what our rules say, word for word:
Every agent is /clear'd before each task and runs as a stateless executor.
Durable state lives ONLY in the TMS, the LLM-wiki, and the twin, never in a session.
Read the wiki before you plan; write your findings back when you finish.
This has a cost: an agent has to re-read context. It also has a benefit. When one agent's session goes wrong, the next task starts clean, and nothing important was only in that agent's head.
Where the loops come in
A company is a graph of work, and a lot of that graph is repetition: check a thing, report it, act if it changed. For that part we use bounded loops. A loop has a producer, a cadence, a turn cap, a time limit and a stop signal. In agent.ceo that pattern runs today: a loop can be created, its output read, and a stop signal sent, after which one full cadence passes with no new output.
Rendering diagram…
What we got wrong first
The rules exist because each one replaced a habit that failed. We treated "the pull request merged" as delivery, and pages stayed stale. We let agents mark their own work complete, and the checks they ran passed on broken work too. We read an unchanging dashboard as calm, when the thing feeding it had stopped.
None of that is special to AI. Human teams make the same mistakes. The difference is that agents make them faster and without complaint, so the rules must be written down and loaded every time.
The method behind the verification rules is in the FOVEA loop. The one we lean on hardest is covered in should an AI agent verify its own work?
FAQ
What does it look like when a company runs on AI agents?
At agent.ceo it looks like a set of written operating rules that every agent loads before it starts work. The ones that matter most: no agent certifies its own work, a blocked task must name who can clear it, no progress is treated as an alarm rather than as stability, and a short list of decisions always goes to a human founder.
Which decisions should stay with a human when AI agents run a company?
We keep four with the founder: credentials and secrets, financial commitments, accepting a security risk, and strategic pivots. Everything else is decided inside the organization, and a second party checks the work instead of a human approving it.
How do you stop AI agents from repeating a failing action?
With a written limit. Our rule is that after the same action fails five times, the agent stops and either breaks the problem down or escalates it. The fifth attempt is treated as evidence that the approach is wrong, not as a reason for a sixth.
Related: What is the simplest loop pattern for an AI agent? · Where should a human approve in an AI agent workflow? · Verification as code