The company is the unit

The common question is "how do I get an agent to do this task well." It is the wrong altitude. A task done well is a tool. The leverage sits one level up: a standing structure that runs a whole portfolio of work on its own and asks you only to decide. Call it a company.

The thesis

This site already argues that knowledge management is not filing, it is compilation: turning captured noise into durable structure. The factory is the sequel. If a knowledge system gives Claude a brain, the factory gives it a body. The brain compiles what is known. The body operates on it: it builds, publishes, sells, measures, and repairs. Same operator, one layer of leverage higher.

Two senses of "agent company"

People mean two different things by the phrase, and conflating them wastes time.

• Agents as the product. The agents serve your customers. The company sells agent labor.
• Agents as the workforce. The agents run your projects. They are the staff, not the product.

This is the second one. The output is not an agent. The output is a portfolio of ordinary products (websites, directories, content) that happen to be operated by agents instead of by me.

What I am building

One mini-company that runs my side projects. It has its own wiki layer for memory and state, and the project repositories sit beneath it. It runs on a schedule without me. I read a daily brief, I approve or reject a short list of decisions, and the rest happens on its own, including the part where it checks its own work and rolls back when the check fails.

The split is the whole point. I see everything. I decide the things that matter. I am not in the loop for anything else.

Why the company is the right unit

The honest reason is that the bottleneck was never capability. The scripts to scrape, write, send, and measure already exist and already work. The bottleneck is me sitting in the middle of the loop: remembering what is next, running the right script, carrying the result to the next one, judging it. The company moves that orchestration out of my head and into a system. My role changes from operator to board.

That is also why this is low risk. Nothing here is a new superpower. It is a control plane, a memory, and a scheduler placed on top of work that is already proven by hand.

The honest read

This is an under-proven frontier. The public examples that hold up prove amplification of a capable operator, not autonomy as magic. The rest is mostly demos. So I am treating my own side projects as the testbed: low stakes, real complexity, full ownership. If the pattern cannot run a directory business I understand completely, it has no business running anything larger. That test is the point.