We are pleased to share a recent Design News article authored by Dr Viral B.Shah, CEO and co-founder of JuliaHub and co-creator of the Julia programming language. In this thought-provoking piece, Dr. Shah argues that the next wave of engineering transformation won’t come from incremental automation but rather from Agentic AI that understands physical systems.
In recent years, software engineering has seen an AI revolution of sorts where tools like GitHub Copilot and Claude Code have rapidly moved from being autocomplete assistants to collaborators. Unlike software engineering, industrial or hardware engineering has remained time intensive and slow. Designing physically consistent and accurate models still needs weeks of effort right from assembling components, debugging equations to validating simulations and tracing parameter data across diverse sources. Where software thrives on symbolic, modular artifacts with shorter feedback cycles and tighter delivery, hardware engineering remains restricted because of the complexity of interpretation, intricate systems, and costly validation.
The article challenges the notion that hardware has not made the same leap as software because of the difference between physical and abstract domains. The main contention is the nature and structure of the work itself: software engineering benefits from rapid iteration and reversible errors while hardware engineering is characterized by irreversible, expensive validation and risk-averse processes. It is this fundamental difference that limits throughput and exploration.
The question then is how can hardware engineering reap the same benefits as software development, leveraging agentic workflows and AI. The article emphasizes how real progress in hardware AI needs systems that follow the same philosophy, models, constraints and physics, not simply numerical execution. Dr.Shah takes the example of a friction brake system in an electric vehicle to show the many stages from defining equations to ensuring energy conservation, that typically slow down the engineering. But an AI that systematically searches for and validates parameter values, lists assumptions and documents model choices, accelerates these phases dramatically without compromising on human judgment.
At the core of this argument is a shift from assistance to partnership: agentic AI systems should treat models as first-class objects, not opaque code, letting engineers focus on judgment, tradeoffs, and design insight, while agents handle setup, exploration, and verification.
Dr. Shah also explains why this paradigm requires more than large language models alone. To be trustworthy in engineering contexts, AI must work with machine-interpretable representations of physical systems where conservation laws, units, and explicit assumptions are enforced.
Platforms like Dyad demonstrate this approach by enabling agentic interaction over structured models. Agents can construct physically consistent models from intent, lay down the physical constraints, run validated simulations, and present results aligned with engineering reasoning, effectively shifting engineering workflows toward deeper exploration, better decisions, and more robust hardware systems.
