Why Roofus, and how we'll build it

Context

“Agentic AI” means an AI model that works in a loop: it decides on an action, uses a tool, looks at what happened and decides again, until a job is done. It is the difference between a chatbot that answers questions and an assistant that can actually file the invoice.

Simon wants to understand it well enough to run a business with it, but does not plan to write the code by hand.

Options

1. Learn the theory first, build later. Safe, slow, and abstract.
2. Hand-code everything to learn properly. Thorough, but trains the wrong skill. When agents build, the scarce skills are deciding what to build, specifying it clearly and judging the result.
3. Direct agents from day one, and learn by reviewing what they do.

Decision

Option 3. Simon acts as director and product owner. Claude acts as trainer, planner and reviewer, and sends build agents off to do the work. Each piece of work starts as a written task in the repo and ends as a reviewed pull request.

Consequences

• Tasks have to be written well enough that an agent can build them without asking questions. Writing them well is the first skill to learn.
• Agents sound equally confident whether they succeeded or not, so nothing merges without automated checks and a human sign-off.
• Agents forget everything between sessions. This journal and the repo’s CLAUDE.md are how they remember.
• We start with one agent at a time. Running several in parallel waits until the foundations and conventions exist, or they will make conflicting choices.