Jun 25, 2026

Semiconductor Engineering convened a panel of industry executives at the 2026 ESD Alliance Executive Outlook meeting to examine the benefits and drawbacks of agentic AI in semiconductor design and verification, as reported by the publication.

Agentic AI Reshaping Design Flows

Cindy Cui of ChipAgents stated that agentic AI is already transforming the entire design flow, with applications ranging from design verification and RTL generation to UVM and formal verification, spanning frontend to backend tasks. She noted that the technology is mature and helping engineers accelerate design cycles, but the primary challenge lies in organizational transformation—educating engineers to adopt new solutions and fostering teamwork across the industry.

Ann Wu of Silimate described her company’s focus on building AI models and agent harnesses that act as co-pilots for chip designers, enabling a more directed and massive search of the design space. She pointed out that previous chip design generations were constrained by manual and compute limitations, whereas now engineers can generate many more experiments. A key challenge, she added, is determining the quality of those experiments quickly and accurately.

Shelly Henry of Moores Lab AI, speaking from a user perspective, questioned why engineers still take two years to build a chip despite advanced tools from major vendors. She suggested AI could compress that timeline to three months by generating designs, verification collateral, and performing debug. However, she emphasized the need for proof that generated RTL code, specifications, and testbenches are correct, and that coverage closure does not inadvertently exclude items.

Access to Fabrication and Ecosystem Dynamics

Henry also addressed the difficulty for small companies and startups to secure shuttle capacity at advanced nodes, noting that hyperscalers such as Google, Microsoft, Meta, Intel, and AMD absorb most fab capacity. She expressed belief that AI technology will democratize chip design, enabling more entities to build chips cheaper and faster, which could foster a broader fab ecosystem.

Trust and Human Oversight

Vince Wong of Verific highlighted that while agentic AI will boost productivity, there is a risk of excessive trust in AI systems. He cautioned that AI is not yet ready for fully autonomous operation, and many design flow steps still require fixed points of human interaction. He argued that this need for human oversight should be integrated into every methodology.

Wally Rhines of Silvaco pointed out that the process side of chip development has been overlooked. He noted that the complexity and time required for prototype wafers make traditional physical prototyping impractical, and surrogate models for unit processes could enable process integration on a computer. He described this as a significant opportunity, given the high value placed on earlier ramp-up times by companies investing billions in wafer fabs.

Integration and New Flows

Rhines added that many in the EDA industry may end up selling or renting agents, or providing databases to customers who wish to keep proprietary information confidential. He suggested that agents could work with customer data to calibrate synthetic data against predictive norms.

Dave Kelf of Breker Verification Systems reflected on the industry’s historical pattern of worrying about keeping up with chip design progress. He stated that AI will make a dramatic difference, but the question is how to achieve that. He noted that while some companies are incrementally building AI into existing tools, there is also an opportunity to create entirely new flows, though such radical changes have often failed in EDA. He stressed that chip design is an exact science where even a 1% error is unacceptable, so ensuring AI devices are exacting is a critical problem to solve.

Measuring Success and Quality

Rhines emphasized that success with agentic AI is quality-oriented rather than quantity-oriented. He said it will reduce design time, catch more bugs, and eliminate drudgery in documentation, RTL generation, and testbench creation, freeing people for more creative work. The quality of intermediate and final designs, he argued, is what differentiates agentic AI from traditional methods.

Cui described metrics as an evolving system. She noted that while past AI helped individual engineers speed up, agentic AI can accelerate entire engineering organizations with trust, which she called trusted acceleration. She envisioned a future self-evolving engineering system that learns from past experience and continuously improves.

Wu stated that metrics will evolve alongside system capabilities, but outcomes must remain anchored. She referenced the software industry as a leading indicator, where the focus shifted from AI adoption rates to return on investment and efficiency. For her, the litmus test is whether agentic AI enables things previously impossible, such as compressing timelines from 12–18 months down to 6–9 months through hyper-efficient processes. She stressed the need to ascertain quality quickly during exploration and validation.

Henry drew an analogy to a math prodigy who could multiply large numbers quickly but whose speed was useless without accuracy. She noted that engineers vary in skill, and trust in their output is currently managed through ad hoc checkpoints and proof points that differ by company. She called for community consensus on benchmarks and quality standards that all agents could strive toward.

Kelf mentioned that industry players are cooperating on benchmarks to measure improvement from collaboration. He described this as a pie-growing exercise where cooperation can make a big difference, whereas competition could cause the effort to fall apart. Establishing metrics and understanding flows, he said, is critical to success.