Quand: le vendredi 22 mai à 10h00

Où: C06, Evry

Title: Agentic AI in Software Engineering: From Automation Gains to Reliability Gaps

Abstract:

Software systems are among the most complex artifacts ever created by humans, featuring millions of components interacting between each other, continuously evolving under tight constraints of quality, time, and scale. The recent emergence of Agentic AI systems powered by Large Language Models (LLMs), such as Copilot, Cursor, Claude Code, Devin, and Codex, marks a fundamental shift in software engineering practice. These systems move beyond assistive automation to autonomously author code, submit pull requests, trigger CI workflows, and iteratively repair failures, effectively acting as software engineering agents rather than tools. While early evidence demonstrates substantial productivity gains, recent empirical studies reveal a more nuanced reality. Agentic AI introduces new reliability challenges, including increased CI failures, reverted changes, self-admitted technical debt, and a persistent need for human oversight and intervention. Rather than eliminating human effort, Agentic AI reshapes it, shifting developers’ roles from direct implementation toward supervision, guidance, validation, and quality control.

In this talk, I synthesize recent empirical findings on Agentic AI in software engineering, covering agent-authored pull requests, CI reliability, failure recovery, technical debt, and human-AI collaboration patterns. I will discuss where Agentic AI excels, where it struggles, and why failures often stem from context misalignment, scope management, and infrastructure constraints rather than code generation quality alone.

Short Bio. Ali Ouni is a Full Professor at École de technologie supérieure (ÉTS Montreal), where he leads the Software Technology and Intelligence Research Group (STIL). He is a passionate software engineering researcher and educator. He has developed pioneering research work in the area of software engineering, software maintenance and evolution, software quality, and empirical software engineering. He leverages advanced artificial intelligence techniques to address challenges related to software products, processes, and stakeholders. He published over 170 peer-reviewed articles many of which has repeatedly appeared in top venues in software engineering and received several Best/Distinguished Paper awards. His research work has been done in collaboration with major industrial software companies. He is the recipient of the Outstanding Early Career Computer Science Researcher Award from the Canadian Society of Computer Science CS-Can/Info-Can in 2024, the Research Excellence Award (Prix Relève) from the Université du Quebec in 2023. He obtained his PhD degree from the Université de Montréal, with the Excellence Award. He is also an area editor of the Software Engineering Body of Knowledge (SWEBOK) book, version v4 (2024), published by the IEEE Computer Society. He is an IEEE Senior Member.