The fleet: companies run by agents

Most of the engineering in my companies is done by AI agents. Not assisted by them, done by them: a fleet of workers that take a task, branch, build, test, review each other, and bring the result back for a decision. I supervise. I decide what to build and whether what came back is right. The hands on the keyboard are, for the most part, no longer mine.

I want to describe this plainly because it is easy to describe it badly. It is not magic and it is not a trick. It is closer to running an engineering organisation than to writing code, with the usual problems of an organisation: unclear briefs produce confident nonsense, work that nobody reviews drifts, and the loudest signal is not always the right one. The craft moved up a level. I now spend my days on intent, on specification, on review, and on the design of the system that does the work.

A fleet works the way a river works at the head of its delta. One task enters and is divided, and divided again, until every channel carries exactly what one worker can use. A single large task is useless to a single agent; divided finely enough, with the right structure between the channels, it becomes a day’s work for forty of them.

What I have learned so far, in the form of rules I keep:

• The brief is the product. A fleet amplifies whatever you put in, including your vagueness.
• Review is not optional. Every branch returns to the trunk through a gate, and the gate is where quality lives.
• Isolation beats coordination. Workers that cannot touch each other’s work cannot ruin it.
• Keep the person in the loop at the decisions, not at the keystrokes.

This is also the subject on which I feel most responsible to speak carefully. A fleet that writes software for a company that serves physicians is two steps from a patient. The rule from the essay on what a machine must never decide applies here as much as in the clinic: the machine may do a great deal, and it must not be the one that decides what matters.