Can AI Design with Care? Evaluating an AI-Based Multi-Agent System for Beginner Smart Home Projects
Stephen Yuan
21/07/2026
Large language models are increasingly used by students to brainstorm and explain engineering projects, but a fluent design proposal can still hide weak feasibility judgment, incomplete safety reasoning, or shallow attention to human needs. This student-led project asks whether AI can design with care by evaluating a Gemini-based multi-agent system for beginner smart home projects. The system uses role-based agents, including a facilitator, planner, engineer, and safety reviewer, to break a smart-home idea into requirements, parts, wiring, software logic, testing steps, and risk mitigation. The study evaluates 40 generated proposals using a human-scored rubric and written comments. Across the 40 prompts, the overall average score was 4.375 out of 5. The system performed especially well on completeness and readability, with mean scores of 4.950 and 4.925. It scored lower on ethical considerations and care for vulnerable users, with means of 3.700 and 3.650. The comments show that the system often sounds caring and safety-aware, but it sometimes overestimates beginner feasibility, treats warnings as if they solve dangerous design choices, and inconsistently recognizes accessibility benefits such as manual override, reduced physical effort, and privacy-preserving sensors. These findings suggest that multi-agent AI systems can be useful early-stage engineering tutors, but only when students and teachers keep asking whether a complete-looking design is actually safe, realistic, and humane.