§1.5 From the Engineering Problems to MAGE

The preceding discussion suggests a different unit of analysis for agentic software engineering. A coding agent operates within an environment that supplies system knowledge, instructions, tools, and evidence, and that may constrain its actions or reject its results. The engineer can develop and maintain this environment without controlling the foundation model’s internal representations or inspecting every line of implementation. Engineering responsibility still includes determining how the system ought to behave and how it ought to be constructed. The larger engineered system provides the means to carry those decisions into autonomous work, direct that work, and evaluate its results.

Modeling and Alignment address the two engineering problems in complementary ways. Modeling makes consequential properties explicit in representations suited to the questions being asked, reducing the system knowledge that people and agents must reconstruct during each task. Providing a model, however, does not ensure that its obligations will be respected. A coding agent may overlook a constraint, interpret it incorrectly, or violate it while satisfying another requirement. Alignment connects selected obligations to mechanisms outside that producing judgment: constraints restrict actions or states, sensors produce evidence, and gates evaluate evidence and control whether work may proceed. Figure 1.5-1 summarizes these relationships. The principles can also serve purposes beyond their primary motivations here. A model can improve assurance by making a property checkable, while a gate can enforce a permission without an elaborate system model. Engineers can therefore use either independently and combine them where the work warrants both.

From agentic machinery to MAGE Two rows connect substrate properties to engineering problems and MAGE responses: finite working state to the Representation Problem and Modeling; probabilistic agentic work to the Enforcement Problem and Alignment. Finite working state Representation Problem Modeling Probabilistic agentic work Enforcement Problem Alignment SUBSTRATE PROBLEM RESPONSE
Figure 1.5-1. From agentic machinery to MAGE. Finite working state motivates Modeling; probabilistic autonomous work motivates Alignment. The two principles address different engineering problems and operate together.

I will use DocAble to illustrate how the two principles meet in practice. Its remediation pipeline must improve document accessibility while preserving the original content. A representation of extracted content makes it possible to compare the original document with the result. The validator checks whether the revised document has lost any of that extracted content. The completion rule fails the job and refuses the result if this check fails. The representation and validator perform narrower jobs than remediation itself: the agent still proposes changes, and the engineer must determine what accessible content should convey. Together, the representation, check, and acceptance rule preserve a consequential engineering decision across independently produced changes. The obligation can remain in force without requiring each implementing agent to remember it or certify that its own work complies.

This arrangement concentrates engineering judgment on the properties, evidence, and freedoms that matter. The engineer must decide which properties deserve explicit representation, which obligations require independent enforcement, and which implementation choices can remain open. Those decisions depend on the system’s purpose and the consequences of getting them wrong. A model that records every implementation detail may be expensive to maintain and unnecessarily restrict the agent. A check over an inadequate representation may reliably establish the wrong property. Conversely, leaving a consequential requirement implicit forces each new agent or engineering episode to reconstruct it. MAGE asks engineers to make these choices deliberately: preserve the distinctions and obligations that justify durable structure, provide evidence suited to their consequences, and leave genuinely acceptable alternatives open to the implementer.

The same engineering questions arise outside software, although the representations and evidence differ. Mechanical engineers work with geometry, tolerances, simulations, and physical tests; accountants use ledgers, reconciliations, and audit evidence. The appropriate balance between delegated work and explicit governance depends on the reasoner’s capabilities and the domain’s obligations. Chapters 2 and 3 develop Modeling and Alignment separately, Chapter 4 shows how they work together as systems evolve, and Chapter 5 examines DocAble and other software factories. Chapter 6 develops the theoretical argument, while Chapter 7 considers its implications beyond software.

© James C. Davis, 2026–present