Abstract:

Generative Artificial Intelligence (GenAI) and LLMs have emerged as a transformative force in many fields, fundamentally redefining how professionals’ approach traditionally complex tasks, including software engineering tasks. An LLM-based agent is an intelligent system that uses an LLM as its central reasoning engine to autonomously address complex tasks such as software engineering tasks. Thus, unlike standard LLM, these agents plan, decompose problems, use external tools, and iteratively evaluate and improve their actions, distinguishing themselves by their ability to reflect and adapt their behavior over time.

The academic literature has documented the impact of AI in various subfields of software engineering, including coding, testing, analysis, patching, and documentation. The concept of vibe coding, that is, the practice of guiding LLM-based agents through natural language to generate code by describing the objective or "feel" of the desired outcome rather than specifying each logical step, represents a clear manifestation of this shift towards agentic AI in software development. This evolution transcends the simple use of code assistants to encompass specialized agents across multiple software engineering domains: from automated requirement scraping to intelligent test generation, advanced debugging, software quality improvement, and even automated reverse engineering.

This revolution has been particularly intense in educational contexts, where software engineering students find themselves navigating a transition from traditional algorithmic approaches to human-agent collaboration paradigms that demand new competencies and working methodologies.

However, the advent of LLM-based agents has introduced an additional layer of complexity and opportunities. The autonomy in deploying tools with LLM backends grants agents tremendous power and potential for automating software engineering tasks, but this increased automation does not immediately translate into productivity gains, particularly in educational contexts where trust in agent outputs and manual effort required to integrate agentic workflows constitute critical challenges.

In the specific domain of higher education in software engineering, this transformation acquires particular dimensions. Students must not only master fundamental technical concepts but also develop competencies in effective prompting, that is, the fundamental ability to interact with LLM-based agents where output quality depends crucially on instruction formulation. Additionally, they must navigate the ethical and pedagogical implications of integrating these agents into their learning process, from academic integrity concerns to the impact on their creative development and cognitive load.

Following the methodological tradition established in our previous work on grounded theory in software engineering, this study adopts a qualitative approach to examine how students in software engineering adapt their work and learning processes in response to the emergence of generative agents. Our objective transcends mere tool cataloging to explore the student-agent relationship in key software engineering tasks, aiming to propose curricular design guidelines that integrate agents with trust guardrails ethically and assessments adapted to the new times of learning.

Keywords:

Agentic AI, Software Engineering, Higher Education, Qualitative Study.