Model-Based Agentic Software Engineering

How should we engineer software when implementation becomes abundant but engineering judgment remains scarce?

A single centre spine reads top to bottom: Scale and Finite reasoning combined in one node, a New engineering imbalance, the Modeling Principle, the Alignment Principle, the Governed Engineering Environment, and Engineering Capital. Scale is the enduring source of a finite reasoning horizon: large systems exceed any one reasoner, so relevant state must be abstracted or reconstructed, which makes judgment scarce at scale. Commodity intelligence enters laterally at the New Engineering Imbalance, making implementation abundant while judgment, representation, evidence, and authority stay scarce; it changes the economics of the imbalance but does not cause the reasoning horizon. The Modeling Principle answers the imbalance by making consequential knowledge and intent explicit at useful scales, which exposes those properties for analysis and reaches the Alignment Principle, where selected obligations are given authority through constraints, sensors, validators, and gates. Passing through Alignment the spine enters the Governed Engineering Environment — the synthesis of both principles — and effort compounds as Engineering Capital when future value exceeds its carrying cost. A single subordinate governance-conversion loop feeds recurring failures and judgment back to adapt the models and controls, strengthening the environment. Two dashed failure paths lead to Engineering Churn: weak representation is repaid as reconstruction, weak authority as repeated adjudication. A light dotted long-range rail marks capital reopening the reasoning problem as scope grows.

Capable agents have made it much cheaper to produce working code. They have not made it cheaper to decide what should be built, which obligations govern it, how to read the resulting evidence, or whether the system is acceptable to ship.

MAGE studies the engineering structures that make delegated implementation governable: making consequential knowledge explicit as models, giving selected obligations authority over what agents produce, validating realizations against those models, and turning recurring human judgment into structures that later work can reuse.

The engineering problem

Capable agents have made producing working code much cheaper. They have not made it cheaper to decide what should be built, which obligations govern it, or whether the result is acceptable.

• Model-Based Agentic Software Engineering
  The book-length statement of the framework.
• Software Supply Chains are Dead: Use-Case-Oriented Regeneration
  Argues that when regeneration is cheap, reuse decisions change shape. The position that connects the supply-chain work to the agentic setting.

What agents actually do

Claims about delegated engineering need evidence about how agents behave on real systems.

• SysLLMatic: Large Language Models are Software System Optimizers
  arXiv · 2025
  Showed that language models can improve the measured performance of real software systems, and established the measurement setup the later agent-optimization work builds on.
• How Do Agents Perform Code Optimization? An Empirical Study
  International Mining Software Repositories Confe · 2026
  Characterizes what agents actually do when asked to optimize code, which is the empirical ground for claims about where judgment still has to sit.
• Beyond Local Code Optimization: Multi-Agent Reasoning for Software System Optimization
  JAWs · 2026
  Extends optimization from local edits to reasoning across a system, which is where delegated work starts to need explicit models rather than context.

Infrastructure for governed work

Delegated work needs somewhere to be discovered, verified, and reproduced.

• AgentHub: A Registry for Discoverable, Verifiable, and Reproducible AI Agents
  JAWs · 2026
  A registry for agents, treating discoverability, verification, and reproducibility as properties an agent ecosystem has to provide rather than properties individual users establish.

Funding and support

This work has been supported by:

• US National Science Foundation — CAREER: PTM-SEER: Software Engineering Foundations for Re-Using Pre-Trained Neural Models (#2541917)
• US National Science Foundation — RFE: Research: Developing and Piloting a Prompt Engineering Competency Framework for Software Engineering Education (#2452533)

Explore the MAGE book, course, and framework →