In this work, we investigate the role of generative AI agents as reflective partners in engineering design. While such models are increasingly used to generate design solutions, concerns remain about their potential to diminish designers’ critical thinking and reasoning skills. To address this, we develop a mixed-initiative conversational framework that positions large language and vision–language models as reflective thinking partners rather than solution providers. The framework is structured around five contextual information channels, namely task–role context, design representations, historical context, evaluation signals, and target references, that enable AI agents to ask reflective questions and provide explanations and suggestions. To study this framework within a concrete design context, we develop an interactive tool that embodies the notion of contextual fidelity of 2D structure design tasks. We implement varying levels of contextual fidelity, defined by the extent of contextual information available to the agent. We evaluate these levels through a between-subjects study with forty-six participants, comparing a high-fidelity and a low-fidelity agent against a control group without AI support. We examine the impact of the agents on how users think, talk and act, using a comprehensive set of metrics, including coarse-level design objectives (deformation and material usage), solution quality metrics (structural and geometric analysis), process-oriented measures (design space exploration patterns and trajectories, design strategy shifts), conversational dynamics (thematic and temporal analysis), and subjective surveys (NASA-TLX, Cognitive Load Theory, Trust in AI). Our analyses show that while conversational agents do not immediately help improve coarse-level design objectives, they significantly shape nuanced aspects of design processes and outcomes. Interaction with the agents critically influences how users explore the design space, where agent-supported groups exhibited more focused exploration patterns compared to the control group’s broader trial-and-error approaches. Furthermore, interactions with the high-fidelity agent led to solutions with higher symmetry and topological alignment with optimal designs, fostered deeper reflection, reduced mental demand, and supported more deliberate design decisions. Building on these findings, we discuss broader implications of AI agents for problem-solving processes and outline guidelines for designing adaptive and generalizable frameworks for different domains.

Augmented Reality (AR) training is a cost-effective and safe alternative to traditional instructional methods. However, training novices in basic mid-air AR interactions remains challenging. To address this, we aimed to: (a) develop a robust metric to evaluate user performance across different AR interaction techniques and develop adaptation models to predict additional training requirements; (b) evaluate the adaptation models using a neuroergonomics approach. We conduct a two-phase study during which, novice participants perform simple AR interactions: poking and raycasting. In Phase-I, twenty-seven participants’ data is used to identify a bi-variate performance metric based on median completion ime and consistency. Unsupervised models are trained using this metric to classify participants as low/high performers. In Phase-II, we evaluate the models on twenty-one new participants and analyze the differences in performance, neural activity and heart-rate variability between low/high performers. Our study showcases the effectiveness of our models and further discusses the potential of integrating neuroergonomics for advanced AR-based training applications.