What 40 Years Of Design Automation Can Teach Us About AI Agents

Thomas Andersen is Vice President of AI and Machine Learning at Synopsys.

I still own the graphing calculator I bought as a high school student in the late 1980s. I rarely use it anymore because the software tools I use today offer those same capabilities and much more. But the calculator provides a visible reminder that established tools don’t go away when more advanced technology arrives. Their capabilities are absorbed into newer tools, and the work moves up a level.

That pattern has repeated for four decades in the semiconductor industry.

Around the time I bought that calculator, logic synthesis transformed chip design by allowing engineers to describe what a circuit should do and letting software determine how to implement it. Automated placement and routing later took over much of the physical layout work engineers once performed by hand. Formal verification allowed software to prove that a design satisfied specified requirements.

Each advancement raised a familiar objection: The tool can’t possibly understand the design as well as an engineer does.

Today, many industry leaders are asking how much chip design work and decision-making should be entrusted to AI agents. Based on lessons from 40 years of design automation, here’s what I would tell them.

Trust will ultimately determine adoption.

In every previous wave of automation, the technology was ready well before organizations were. Engineers didn’t stop reviewing synthesis results as soon as the tools became available. They reduced their oversight gradually as the technology proved reliable and its limitations were understood. Trust was earned narrowly, then extended.

The mistake I see today is focusing on what agentic AI can do without defining where it can operate reliably and when human review remains necessary. If teams can’t pinpoint the situations in which an agent’s work should be reviewed, they’ll either check everything, which eliminates the benefit, or check too little, which introduces unnecessary risk.