§ 07 / Factory

The factory

If a knowledge system gives Claude a brain, the factory gives it a body. Notes on running a portfolio of side projects with an agent workforce, while staying the only decision-maker. Written from the system I am building, not in theory.

  1. 01The company is the unitMost agent work optimizes the task. The leverage is one level up: a standing company that runs a portfolio while you only decide.
  2. 02The shape of the companyFive working parts and a wiki. An owner who decides, a lead that plans, departments that do the work, safeguards that gate it, and a memory that ties it together.
  3. 03The wiki is the memoryA scheduled agent remembers nothing from yesterday except what is written down. The wiki is not documentation for the company, it is the company's working memory.
  4. 04The workforce is scripts that reportThe departments are not new intelligence. They are the scripts that already work, wrapped so they run on schedule and write down what they did.
  5. 05The gate is what buys autonomySafeguards are usually framed as the friction that slows you down. Here they are the thing that lets you let go. You can only hand off what you can automatically check.
  6. 06Decisions are executable specsA decision is not a yes or a no. It is the exact action, ready to run, with its evidence and its undo attached. My job becomes ruling on specs, and autonomy is earned one department at a time.
  7. 07How I see it and steer itThe company runs without me, but I am never blind to it. The interface is four surfaces, and none of them is a new app. I read a brief, I rule on a queue, and I talk to the lead.
  8. 08A project is a charter, not a rebuildGoing from one project to many is not multiplying the work, because the departments and the wiki are shared. A new project is a one-page charter pointing at capabilities that already exist.
  9. 09The testbed: a directory portfolioPolish directory sites are the first thing the company runs. This is the setup and the test, not the scorecard. The numbers and the post-mortem come once the loop has run real cycles.
  10. 10FrontiersThe open questions I have not answered. The review-bandwidth ceiling, where earned trust breaks, the pull between cost and trust, and whether the whole thesis holds. This page is living.