Siemens EDA's Harry Foster makes a pointed argument: EDA tool performance is no longer the bottleneck in RTL verification. The constraint is coordination -- engineers spending cycles interpreting results, adjusting strategies, and re-running across multiple tools as designs evolve. Traditional automation breaks here because RTL verification is inherently adaptive, not a fixed pipeline. Agentic AI, operating across workflows rather than within single tools, is the architectural response.
The key distinction Foster draws is between AI that sits outside the tool chain (parsing logs, generating scripts) and AI that is engine-native -- invoking simulations, querying coverage models, analyzing failures through structured interfaces. The former adds overhead. The latter actually closes the loop. This is the same integration vs. wrapper debate playing out across software engineering tooling, but with much higher stakes when sign-off quality determines tapeout readiness.
What is refreshing here is the explicit constraint on autonomy. Foster is clear: full autonomy is not the goal, and it is not desirable. The practical deployments -- lint and static analysis, CDC verification, debug correlation -- use AI to accelerate execution and surface insights, while sign-off authority stays with the engineer. Explicit approval points are an architectural decision, not just a safety hedge. That is the right mental model for any AI-assisted design flow.
The counterpoint worth tracking: this is a Siemens EDA piece, and the specific agent architecture Foster describes maps closely to Siemens' own toolchain. The ideas are sound, but the implementation claim of "engine-native interfaces" will require independent validation. Watch for competitive responses from Synopsys and Cadence -- they are not standing still on agentic verification either.