Skip to content

How do we safely grant autonomy to commodity intelligence?

This is a book about engineering: what engineering becomes when reasoning capacity is abundant. Engineering is the discipline of exercising informed control over consequential systems and accepting responsibility for their outcomes. Engineers need not personally perform every act required to realize a system. They must understand and direct the work, evaluate evidence for the properties that matter, and intervene when necessary. They remain answerable for the consequential decisions made under their authority.

Society has long entrusted consequential work to computers. Software tallies tax returns, flies aircraft, and operates power plants. Traditionally, software engineers translated human intent into programs and controls that computers could execute. Now, general-purpose machine intelligence can interpret intent, make judgments, and produce implementation cheaply enough and at enough scale to become an ordinary engineering resource. In this book, I call this increasingly available capability commodity intelligence 11. John Nosta, “Intelligence as a Commodity,” Psychology Today, March 12, 2026, https://www.psychologytoday.com/us/blog/the-digital-self/202603/intelligence-as-a-commodity. 22. Darwin Ling, “The $0.50 Barrel of Intelligence,” GoodAI, July 20, 2026, https://goodai.substack.com/p/the-050-barrel-of-intelligence., and the tools that supply it, agents.

Agents can be used to produce and modify software at rates far exceeding what human engineers can review. Engineers need ways to preserve control, because engineers remain responsible for ensuring that the resulting systems satisfy their requirements. Important decisions must survive successive changes, and engineers need evidence for properties that may be difficult to see in any one change. How should we engineer the environment in which that autonomous work takes place?

MAGE: Model-Based Agentic Engineering

Agentic engineering is the engineering of systems and environments in which autonomous agents perform consequential work. Model-Based Agentic Engineering (MAGE) is the approach this book develops. It rests on two principles. Modeling makes consequential engineering knowledge and intent explicit in purposeful representations. Alignment makes selected engineering obligations enforceable, so those obligations can hold independently of an agent’s judgment.

The MAGE method combines these principles through four recurring engineering activities:

  1. Model consequential knowledge. Choose the engineering question, then represent the facts, relationships, and decisions needed to answer it without carrying unnecessary detail.
  1. Enforce important obligations. Use constraints to restrict actions or states, or gates to admit work only when required conditions hold.
  1. Do the governed work. Let agents operate within the resulting environment, using its models and mechanisms while exercising the implementation freedom engineers have left open.
  1. Convert recurring judgment. Identify engineering decisions that require repeated human intervention, then encode them in models, procedures, or mechanisms that subsequent work can inherit.

Together, these activities develop a governed engineering environment: representations help humans and agents reason about the system, while mechanisms enforce selected obligations. The activities need not occur in a fixed order, and the environment can change as experience reveals what needs better understanding or control. Useful structure becomes engineering capital when later work benefits from it.

Why MAGE? Why Models?

One way to understand MAGE is as a classical software-engineering move: add a level of indirection.** In software engineering, adding a level of indirection is so commonly a solution that it is jokingly called the Fundamental Theorem of Software Engineering. Conventional engineering has long separated the engineering of an artifact from its physical realization. The familiar inscription “Designed by Apple in California, Made in China” is an everyday example: the people who establish a product’s architecture, requirements, and tolerances need not be the people—or machines—that realize it. Drawings, specifications, and other representations preserve consequential decisions across that separation.

When software engineers wrote most of the code themselves, they could make design decisions and implement them directly. Much of their knowledge remained implicit in the implementation and in their minds. That arrangement of work is less effective with agents. An agent needs to know which dependencies are permitted, who owns a resource, and what behaviors must survive changes. Engineering models give that knowledge an explicit form that engineers and agents can inspect and reason through, without repeatedly reconstructing it from the code.

The purpose of this indirection is to preserve important engineering decisions while leaving other implementation choices open. Agents can search, implement, revise, and improve solutions within that freedom. Engineers can impose additional constraints, but more structure is not automatically valuable. The engineer must decide what matters enough to make explicit and whether the resulting understanding, assurance, or efficiency justifies the cost.

The origins of MAGE

MAGE grew out of an engineering problem I encountered while building DocAble, a service for making documents accessible to people with disabilities. I built DocAble by myself—just me and Claude, which produces software far faster than I can inspect. Yet DocAble has to satisfy legal accessibility requirements. I could not solve that problem by supervising the agents more closely; human attention was precisely the resource I did not have.

As I watched the collapse of my weekly Apple Watch sleep score, I realized I had to make the requirements and other consequential properties explicit enough that Claude could preserve and evaluate them without requiring me to inspect every step. I needed to capture my engineering judgment in order to reduce how often I had to supply it.

DocAble provides the book’s deepest worked example: a system developed over several months and used in production at Purdue. I also reconstruct published industrial accounts to examine related mechanisms in other organizations. DocAble provides depth about engineering decisions and their consequences; the industrial accounts provide breadth across production environments. Chapter 5 presents both sources of evidence and their limitations. Chapter 6 develops the theoretical account, and Chapter 7 asks what may transfer beyond software.

Who this book is for

This book is for software engineers who want to understand what changes when implementation can be delegated to capable agents. It is also for architects, technical leads, and engineering managers responsible for the environment around that work.

The book is also intended for advanced undergraduate and graduate students. It assumes familiarity with ordinary software-development practice: source control, testing, APIs, build and deployment workflows, and basic software architecture. No prior experience with coding agents, formal methods, model-based engineering, or machine learning is required. Concepts from those areas are introduced where the argument needs them.

MAGE draws on established engineering disciplines without attempting to replace their textbooks. The machinery described here is a repertoire, not a required architecture. Different problems call for different representations and controls, and some need very little of either. Your task is to judge what structure is worth building, which choices can remain with the agents, and what evidence is sufficient to accept their work.

Works Cited

  1. Nosta, John. “Intelligence as a Commodity.” Psychology Today, March 12, 2026. https://www.psychologytoday.com/us/blog/the-digital-self/202603/intelligence-as-a-commodity.
  2. Ling, Darwin. “The $0.50 Barrel of Intelligence.” GoodAI, July 20, 2026. https://goodai.substack.com/p/the-050-barrel-of-intelligence.
© James C. Davis, 2026–present