A software project built with coding agents moved at high speed, but its quality degraded as the system grew, according to the paper. During the 20-week build, six to eight agents worked in parallel and sustained about 200 commits a day and 1,000 a week.
That tension sits at the heart of an arXiv preprint's question: how can autonomous software work be governed at scale while important engineering properties are made explicit, testable and subject to control, without repeated reconstruction? The proposed framework, MAGE, pairs Modeling, which makes consequential knowledge and intent explicit in task-relevant representations, with Alignment, which gives selected obligations, the requirements a system is expected to honor, authority through constraints, sensors, validators and gates.
The project grew a second layer
DocAble, the longitudinal case at the center of the analysis, was developed with coding agents as its primary implementation workforce. It reached about 540,000 lines of production code and 1.6 million lines of supporting infrastructure for the governed engineering environment.
The paper calls that surrounding layer a support apparatus. Relative to production source, it grew from 0.85 times the production source after the prototype to about three times the production source in mature snapshots. It peaked at 3.68 times the production source during hardening.
Hardening also left concrete traces in the repository. Lint files designed to catch recurring problems rose from zero to 747, while gate scripts rose from zero to 102. A total of 208 commits paired a fix with a recurrence-catching lint.
Derived checks later found six instances of a previously identified model-code drift class, with no observed recurrence across 56 subsequent feature implementations. Across a nine-stage modeling sequence, the proportion of implementation elements that remained unmodeled fell from 56% to 7.89%. These measures describe change within DocAble; they do not establish a general effect across software projects.
A pattern beyond one repository
The paper's exploratory, interpretive design also includes a purposive comparison of first-party accounts from Cloudflare, Spotify, Shopify, Docker, Siemens and Zenseact. Across those accounts, the paper reports four recurring structures: knowledge is externalized, action is bounded through tools or roles, generation is separated from evaluation, and human authority remains when decisions are not adequately mechanized.
That scope matters. The authors say the comparative evidence can support claims about recurrence and variation, theory development, refinement and analytic generalization within stated conditions. It cannot, on its own, provide causal effectiveness claims or population estimates.
The experiment was narrower
The paper also reports a preliminary two-week, in-situ experiment inside DocAble. Agents were randomly assigned to receive either explicit instructions for using models or no such instructions. In that experiment, the instructed agents required fewer tokens, turns and elapsed time than the no-instruction group, with the reported difference more pronounced for Sonnet.
The experiment adds a preliminary measurement to the paper's broader theory-building agenda. The authors describe the evidence as supporting theory development, refinement, recurrence and variation within stated scope conditions, rather than causal effectiveness or population estimates.
A proposal for the next test
To make the argument testable, the paper proposes three falsifiable hypotheses. They address representation leverage, failures of representation integrity, and whether explicit models can provide repeatable predicates for selected system-level obligations. In practical terms, the questions are whether the right representation reduces reconstruction work, stays faithful to the system and allows important requirements to be checked repeatedly.
The paper's practical message is a division of authority. Models should make intent and consequential knowledge visible; constraints, sensors, validators and gates should enforce selected obligations; and human authority should remain at decisions that are not adequately mechanized. For software teams using coding agents, that amounts to deciding what must be represented, what can be checked automatically and where a person still has the final say.
The document is an arXiv preprint, version 1, dated 25 August 2026. Its evidence is presented as exploratory theory development, not as a causal effectiveness estimate.
Paper data and sources
Original title: Model-Based Agentic Software Engineering
Authors: James C. Davis, Kelechi Kalu, Huiyun Peng, Parth V. Patil
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text