Running an AI-native company

Our World is a small company. AI agents handle much of our development, research, document preparation, and scheduling. This note is about how to build a company around that.

Who does what

The one thing only people can do is judge value. What counts as good differs from person to person and from company to company. An agent can infer from past records what this company would probably choose. But deciding which values to apply, and taking responsibility for the outcome, stays with people.

So we keep the human job narrow: write down what we value, and check whether the results match it. The more of our values we write down, the more an agent can act on our behalf.

Three layers

We think of the company as three layers.

• Rules and skills — who each agent is and how it works, kept in one place.
• Data — records and documents. People work with them through screens and agents through files, but both are working on the same thing.
• Machines — the workspaces agents run in, borrowed when needed and returned when they are not.

Expertise belongs to places, not to agents. An agent belongs to no department; it takes on a role by reading the rules and guides kept in a given place. What a human organization would call hiring and letting go comes down to which guides an agent is given to read.

An example: the secretary

An agent acting as our secretary manages our internal schedule. There is a single set of data: our founder reads and edits the schedule on screen, and the agent does the same through its tools. Notes from people and notes from the agent go in separate fields. Where written rules alone didn’t keep things consistent, we briefly added an automated check (a hook), though we have since removed it.

For anything involving money or communication outside the company, the agent proposes and a person approves. We plan to gather each day’s approvals into a single review, with large amounts and risky items highlighted on screen.

Open questions

• Once we have employees, where do people fit in this structure?
• As the number of approvals grows, how do we keep people from approving things without reading them?
• The name. We first called this “AI-first management,” but that sounds narrower than what we mean. The current title is provisional, too.