AI CODE-GENERATION PARADIGMS, FROM SINGLE PROMPT TO MULTI-AGENT: AN EXPERIMENTAL STUDY WITH IMPLICATIONS FOR SOFTWARE ENGINEERING EDUCATION

Abstract

Between 2024 and 2026, software engineering absorbed artificial intelligence assistants able to build entire applications autonomously, a transition that reshapes professional practice and computing education and still lacks controlled evaluation. This exploratory case study compares four AI code-generation paradigms arranged as an orchestration ladder: single prompt (P0), agentic harness (P1), multi-agent system (P2), and a multi-agent system with a dedicated security agent (P3). Holding model, sampling and stopping limits fixed, and varying only the paradigm, the four arms receive the same frozen specification and rebuild, from scratch, a full-stack web platform for research-laboratory management. Evaluation covers structural faithfulness of schema and contract, functional correctness through end-to-end acceptance tests, quality, security mapped to CWE and OWASP categories, and cost. Across sixteen initiated runs, four per paradigm, no difference in correctness was detected among the three agentic arms, with medians from 34.5 to 35 acceptance criteria passed out of 48 (Kruskal-Wallis among P1, P2 and P3 with a p-value of 0.73), whereas the single prompt was fragile: of four initiated runs, one produced an executable application, two produced non-executable artifacts and one produced no assessable artifact. Correctness did not scale with generation effort and, given the small number of replicates, the entire reading is descriptive. Reliability, not correctness, separates the paradigms. In billed dollars the single-session harness (P1) is the most expensive, due to reprocessing a large volume of cached context. Application security is null across all arms, the only recurring high-severity finding being running the container as root. The study argues that asking for “secure code” in the prompt does not by itself reduce vulnerability, which repositions the role of verification in software engineering education.