Agentic AI tackles RTL verification’s productivity gap

For decades, gains in electronic design automation (EDA) came from faster engines. Verification teams relied on more powerful simulators, scalable formal tools, and higher-capacity solvers to keep pace as designs grew. That approach is now hitting its limit.

In modern system-on-chip development, the real bottleneck is no longer raw tool performance; it is coordination. Verification has become a continuous, adaptive process. Engineers must interpret results, refine intent, and adjust strategies across multiple tools and iterations. Increasingly, productivity is constrained by workflow complexity, not compute capacity .

Regional pressure in Asia

That constraint is especially visible across East and Southeast Asia, where semiconductor investment is accelerating while the supply of specialised engineers struggles to keep pace. China’s chip industry has expanded rapidly as the country pushes for greater self-sufficiency and more advanced IC design and manufacturing. At the same time, rising chip complexity increases the demand for experienced design, verification, and manufacturing talent; public analyses have placed China’s semiconductor talent gap in the hundreds of thousands.

The same talent challenge is shaping policy across Southeast Asia. Malaysia’s National Semiconductor Strategy aims to move the country further into IC design, advanced packaging, and semiconductor equipment while training and upskilling 60,000 high-skilled local engineers. Vietnam has approved a semiconductor human-resources program targeting at least 50,000 workers with bachelor’s degrees or higher by 2030, including engineers across chip design, packaging, testing, and manufacturing.

These initiatives underscore the productivity pressure facing verification teams globally. Talent pipelines take years to mature, but design complexity is increasing now. For engineering organisations, the practical question is how to amplify scarce expertise without removing human judgment from sign-off-critical decisions.

From automation to workflow intelligence

Traditional automation assumes stable inputs and predictable flows. RTL verification doesn’t work that way. Designs evolve, specifications shift, and intermediate results often reveal new risks. This is where agentic AI is gaining traction.

Instead of optimising individual steps in silos, agentic systems operate across workflows. Agents observe the verification state, plan bounded actions, execute tasks, and summarise outcomes.

Integration matters. AI that sits outside the tool chain, parsing logs or generating scripts, can increase review overhead and reduce confidence. By contrast, tightly integrated approaches allow AI-generated actions to be evaluated using the same coverage models, semantics, and checks required for sign-off.

Keep humans in the loop

Despite rapid advances in AI, full autonomy is not realistic for RTL verification. It is also not desirable.

The goal is not to replace engineers, but to reduce manual coordination so they can focus on decision-making and mitigating risks.

Verification decisions often depend on incomplete specifications, implicit assumptions, and trade-offs that require human experience. Determining when results are “good enough” is rarely binary.

That is why effective agentic systems are designed to be human-centred. Agents assist by proposing actions, running tasks, and surfacing insights. Engineers retain control over all decisions that affect intent, scope, and closure.

In practice, this means establishing explicit approval points in the workflow. AI accelerates execution and analysis, but sign-off authority stays with the engineer.

A foundation for trusted workflows

Delivering on this model requires more than adding AI features to existing tools; it depends on an architecture that exposes verification engines in a structured, semantically-meaningful way.

Engine-native interfaces allow agents to invoke tools, retrieve results, and observe system states directly. Instead of relying on unstructured inputs, these interfaces provide controlled entry points, such as running simulations, querying coverage, and analysing failures.

Equally important is context. By maintaining relationships between designs, testbenches, assertions, and historical results, agentic systems can reason across iterations. That continuity is critical in complex verification environments where each step builds on prior outcomes.

Where agentic AI shows value today

Early agentic AI deployments are already improving productivity across several RTL workflows:

• RTL development: AI-assisted code generation can align design intent with verification requirements and catch issues earlier.
• Lint and static analysis: Smarter configuration and contextual filtering help reduce noise and focus attention on real problems.
• Clock domain crossing: Iterative analysis and refinement speed convergence on robust asynchronous designs.
• Verification planning: Agents can translate evolving specifications into structured plans that adapt as designs change.
• Debug: Correlating waveforms, assertions, and logs across runs surfaces likely root causes faster.

Across these use cases, a consistent pattern emerges: bounded automation, strong context, and human oversight. The result is measurable productivity gains without compromising rigour.

Managing risk and maintaining trust

Like any automation, agentic AI introduces new risks. Poorly constrained systems could propagate errors or obscure intent. The solution is not to limit capability, but to enforce structure.

Bounded actions restrict what agents can do. Engine-qualified validation ensures outputs meet production standards. Structured interfaces reduce ambiguity. Mandatory human review preserves accountability at key decision points.

This aligns with the realities of verification, where responsibility ultimately sits with engineering teams. In this model, AI augments expertise rather than replacing it.

Looking ahead

Agentic AI in EDA is still evolving, but the direction is clear. Today’s focus is on reducing friction in planning, execution, and analysis while keeping engineers firmly in control.

Longer term, broader orchestration may emerge, but only with transparency and robust validation.

As design complexity continues to grow, the biggest gains will come from improving how work gets done, not just how fast tools run. Agentic AI represents a shift toward workflow intelligence, helping teams move faster while preserving the rigour required for sign-off.

For verification teams facing increasing iteration overhead, that shift can’t come soon enough.

For a deeper dive into agentic AI, please see my new paper Human-centered agentic AI workflows for RTL verification.