Agentic AI comes to chip design to ease the engineer shortage, not to replace them

By Zendoric · July 22, 2026.

For decades, the electronic design automation industry (EDA, the software tools used to design and verify chips) improved its productivity with faster engines: more powerful simulators, more scalable formal tools, higher-capacity solvers. As argued by Harry Foster, chief verification scientist at Siemens Digital Industries Software, that path has hit a ceiling. The problem is no longer how much compute a tool moves, but the coordination among tools, changing specifications and engineers who must interpret results and adjust strategy with each iteration.

The article, published as an opinion column in Tech Wire Asia, places the problem in a very specific context: the shortage of talent specialized in chip verification (RTL, the level of hardware description —register-transfer level— at which it is checked that a design does what it should before manufacturing). Public analyses cited in the piece put the semiconductor talent gap in China at several hundred thousand. Malaysia has launched its National Semiconductor Strategy to train 60,000 highly qualified engineers; Vietnam has approved a program that aims for 50,000 professionals with at least a university degree in chip design, packaging, testing and manufacturing by 2030. Foster's conclusion: training plans take years to mature, but the complexity of designs is already growing.

His proposal is agentic AI applied to very narrowly defined tasks: agents that observe the state of verification, plan bounded actions, execute them and summarize results, integrated directly into the verification tools themselves (what he calls "engine-native" interfaces) instead of operating from the outside by analyzing logs or generating loose scripts. The use cases he cites —RTL development, static and lint analysis, clock-domain crossing, verification planning and bug debugging— share the same pattern: bounded automation, broad context and the engineer retaining sign-off authority over every relevant decision.

This piece should be read with the right filter: it is a column signed by an executive at one of the world's major EDA software providers, not an independent study. It provides no measured productivity figures, only a conceptual framework and the implicit promotion of an approach that directly benefits Siemens' business. That does not invalidate the argument —the diagnosis of the coordination bottleneck is consistent with what we have been seeing for some time in other domains of agentic AI—, but it does require treating it as the thesis of an interested party, not as verified fact.

That said, the case is a useful counterpoint to our thread on AI and employment. In the administrative back-office, the pressure from agentic AI is substitutive: it automates tasks that people used to do and reduces headcount. Here the opposite happens: the demand for verification engineers far exceeds the supply, so AI's role is not to compete for scarce work, but to stretch talent that is already in short supply. It is the flip side of the same technological phenomenon, and it fits with our underlying thesis: where there is a real shortage of specialized human capacity, AI amplifies rather than displaces.

There is also a structural reason, not just a rhetorical one, why full autonomy is not on the table here. An error in the back-office is fixed with a patch the next day; an error that slips into an already-manufactured chip requires a "respin" —redesigning and re-manufacturing the wafer— that costs months and millions of dollars. In a domain where failure is irreversible and physical, it makes sense that the design of these systems insists on explicit approval points and on keeping sign-off authority human, even as model capability keeps rising. It is a reminder that "how much AI autonomizes" is not only a matter of technical capability, but of how much it costs to be wrong.

In the longer term, if these tools deliver on what they promise, the network effect is no small thing: building AI infrastructure depends, ultimately, on manufacturing more and better chips, and that depends on there being enough verification engineers not to slow the chain. Amplifying that scarce talent with bounded AI is not the flashiest part of the agent revolution, but it is exactly the kind of invisible friction that, once resolved, accelerates everything else.