Coda: Engineering the Factory
The cases in this chapter show several ways that software production is reorganizing around capable agents. Contemporary factories differ in where they place engineering knowledge, how they arrange production, how they divide work between people and agents, and how they evaluate what production yields. The distinction between process design and tolerance gives us a vocabulary for describing those differences without treating them as stages of a common maturity model — factory designers can reasonably place engineering control in different parts of the production system.
The comparison also exposes two costs that become easier to see as realization gets cheaper. Supervision and verification have finite capacity: increasing the rate at which changes can be produced does not automatically increase the rate at which their consequential properties can be established. Engineering structure has carrying cost: the instructions, models, policies, skills, tests, and other structures through which a factory inherits judgment must themselves remain accurate as the software and the factory change. Agent capability changes both tradeoffs, because a better fabricator can reduce the process required around realization without removing the need to establish consequential properties of what it produces.
It also gives us the means to go further. Process design asks how work, information, authority, evidence, and admission are arranged around a capable fabricator. Tolerance asks which variation the factory permits and which properties it must keep under engineering control. Together, those ideas let us model how engineering work is redistributed as realization becomes cheaper and more capable, why some judgments are worth moving into durable production structures, and where the costs of representation, evidence, and human attention should appear. Chapter 6 develops a mathematical model of agentic engineering and a dynamic model of engineering practice, then uses them to derive predictions about how agent capability, engineering structure, and the economics of controlled change interact.