I read Frank Chen's The Company as a Codebase and took notes in a Reading digest. He says every workflow should ship two artifacts: today's result and a versioned improvement to the system that produced it. Foundry sessions close with a writeback pull request into a git-backed company brain. Frontline operators dictate the why behind judgments so an agent can cluster them into doctrine. Autonomy stays inside sandboxes. Humans gate production merges and shared doctrine.
Generated result volume is not the durable company. The result handles today and then depreciates. The company compounds when someone inherits the writeback: a reviewed pull request into ontology descriptions, the prompt builder, and reusable skills. Doctrine goes in that repo. Raw customer records stay out.
Hra is the Codex metaharness that keeps the company loop from vanishing when a window dies. I already named that stack in Building a software factory. Parallel sessions through Hra raise result volume. The writeback into doctrine is the other artifact. Without it, the factory only spends tokens.
Valentine Watkins's Infinite software is not a product says generated code is a category until a hub names live tools. Peter Naur's The theory is the product says carriers are not the theory people hold. Chen's claim sits next to both: even when the tools are named and the theory is held, today's session result still depreciates unless the writeback lands in the system the next session inherits.