How I build

Many agents, one check.

I build with a crew of AI agents. That's how one person runs this many products. It only works with a few rules, and each rule comes from a mistake.

four agentsone checklive
  1. 01

    Draw it first

    Every screen starts as a board in pen.dev with all of its states. I approve the board, then code follows it.

  2. 02

    One agent, one task

    Each agent gets one task, its own copy of the code and a written brief. When a task runs long, it hands off to a fresh agent.

  3. 03

    One check for each batch

    Each agent runs only the tests it touched. When a batch of work is done, it meets in one place and one full check decides.

  4. 04

    One decision at a time

    I don't read agent chat. An agent brings me one decision at a time, in plain words, with the file to open.

  5. 05

    Write it down

    Every mistake becomes a rule, with the story of how it happened.

One rule and its story

An agent waited for the word "passed" and sat for an hour on a check that had already failed. Now every wait reads the result, whatever it says.

More on the blog →