2026 ASEE Annual Conference & Exposition

The integration of Large Language Model-based agents (LLM-based agents), such as GitHub Copilot and ChatGPT, into software engineering practices is advancing at an unprecedented rate. These tools have evolved beyond their original role as information retrieval systems to become active “partners.” This shift has sparked urgent discussions within the field of higher education. While existing research has predominantly focused on the quantitative assessment of the impact of LLM-based agents on coding efficiency and accuracy, their influence on students’ engineering design thinking remains insufficiently explored. Design is the core of engineering practice. Engineering design thinking, which involves skills such as systematic analysis, abstract modeling, trade-off evaluation, and creative problem-solving, serves as one of the foundations of engineering education. Therefore, investigating how LLM-based agents shape students’ engineering design thinking is crucial for understanding and guiding the future trajectory of engineering education.

This study employed a qualitative approach within a semester-long software engineering course at a leading research university. Forty undergraduate students, organized into eight teams, were recruited to develop complex AI-native systems while interacting with LLM-based agents. Data were collected through three methods: (1) post-task semi-structured interviews, (2) interaction logs between students and LLM-based agents, and (3) final software artifacts, including the documentations.

The findings reveal that LLM-based agents took on significant responsibilities for information retrieval and code generation, allowing students to focus more on architectural design. This led to a shift in engineering design thinking process, moving from “ideate-prototype-test” to a rapid “prompt-evaluate-refine” loop. Distinct design strategies emerged. Some students effectively used the agent’s outputs as inspiration, while others exhibited a tendency to “delegate”. Furthermore, cognitive load was redistributed from syntactic implementation to upstream problem scoping and downstream evaluation. The study’s primary contribution lies in providing an empirical foundation and strategic framework for this essential transformation of engineering education.