Many product development tasks require multi-step reasoning and decision-making that exceed the capabilities of Large Language Models (LLMs) operating in isolation. This research introduces a framework for constructing multi-agent systems (MAS) with LLMs for product development and presents an empirical study of its effectiveness. By decomposing workflows into subtasks handled by role-specialized agents, the system leverages agentic design principles and outperforms isolated model outputs. We evaluate this approach using LLaMA-3.1 (8B) and Gemma 2.0 (9B) in a realistic engineering scenario. While enabling more complex problem solving, the MAS also introduces challenges in transparency and fault tracing across input-output chains. Our findings highlight trade-offs and the need for further research on design practices for reliable agentic AI systems in industrial settings.