Human-only vs AI-Assisted Coding: A Controlled Case Study in Agentic Development

Abstract

AI coding assistants have sparked intense debate about their actual impact on developer productivity. This paper presents a controlled case study comparing human-only and AI-assisted coding sessions conducted by a single experienced developer working on the same production-grade full-stack application over consecutive days. We observe an averaged 2.34× increase in lines of code produced per minute when utilizing a single AI agent in a directed, structured pair-programming workflow (10.69 LOC/min human-only vs. 25.03 LOC/min AI-assisted). Beyond raw output, the AI-assisted session delivered full-stack features spanning authentication, team management, project kanban, and a responsive design system — architectural breadth that would have required multiple human-only sessions and skillsets (backend, frontend, UI/UX design, DevOps). We argue that agentic coding, when deployed with clear specifications and human oversight, functions as a genuine force multiplier rather than a code-bloat generator. Further work is needed to establish baselines for multi-agent hierarchies and varied developer experience levels.

Keywords