Electronic Design Automation (EDA) has steadily raised engineering productivity by automating synthesis, optimization, and verification across successive generations of tools. Large language models (LLMs) extend this trajectory by allowing design intent to be translated directly into hardware implementations. In most of the EDA literature, LLM-based solutions are applied only to isolated design stages or individual tasks. That framing obscures the underlying drivers of how capability actually accumulates and how systems scale.
In this paper, we propose three hierarchically related roles that make this accumulation explicit: the Generator, which produces a design artifact in a single generation pass; the Agent, which iterates on tool feedback to refine its output; and the Orchestrator, which coordinates decisions across EDA stages. A survey of published systems reveals a pervasive failure mode we call the syntax trap: models are trained to emit code that looks plausible rather than hardware that is physically correct. Fragmented toolchains and lost design context further hide how a decision at one stage propagates into later stages.
Comparing the three roles shows that current methods still do not scale to industrial designs. What is needed is a standardized, physics-aware orchestrator that connects tools and agents across the full EDA flow, making hardware design both more reliable and more usable.