One-developer case study finds faster code output with an AI agent

What they studied

The authors compared human-only and AI-assisted coding sessions by a single experienced developer working on the same production-grade full-stack application over consecutive days. The AI-assisted workflow used one agent in a directed, structured pair-programming setup, rather than a team of agents working independently.

The comparison focused on lines of code produced per minute. The authors also described the range of application features delivered during the AI-assisted session. This makes the study relevant to builders using an AI coding tool to implement features across several parts of an app.

What they found

The developer averaged 10.69 lines of code per minute without AI and 25.03 with AI. The authors report this as a 2.34-fold increase in code output per minute.

The AI-assisted session delivered features covering authentication, team management, a project kanban, and a responsive design system. The authors argue that this architectural breadth would have required multiple human-only sessions and skills spanning backend, frontend, UI/UX design, and DevOps. That broader comparison is their assessment, rather than a separately quantified result in the abstract.

The authors conclude that an agent can amplify a developer’s capabilities when used with clear specifications and human oversight. They also call for further work involving multi-agent setups and developers with different experience levels.

Why it matters for builders

The useful takeaway is about the workflow as much as the speed: the reported gain came from an experienced developer directing one agent with structure and oversight. It is not evidence that an unattended agent can achieve the same result.

Treat the 2.34-fold figure as a measure of code output, not a promise of finishing an app that much faster. The abstract does not report defect rates, maintainability, or longer-term development outcomes. With only one developer studied, it also leaves open how well the result transfers to beginners.

Source: "Human-only vs AI-Assisted Coding: A Controlled Case Study in Agentic Development" (2026-07-30, Zenodo (CERN European Organization for Nuclear Research)). This is our summary of the paper's abstract — read the original for full details.