A perspective paper from Youngman, Sestito, and Prodromakis names something the field has been dancing around: the reason LLM-assisted RTL generation benchmarks keep improving without producing industrially usable results is that the training objective is wrong. They call it the syntax trap. Models are rewarded for producing code that compiles, not code that closes timing, meets DRC, or survives sign-off. The distinction sounds obvious in retrospect. Most published systems are still optimizing for it.
The paper defines three hierarchical roles. The Generator produces design artifacts in a single forward pass, with no feedback loop. The Agent refines outputs through iterative tool calls, which helps but loses design context at every tool boundary. The Orchestrator coordinates decisions across EDA stages, threading physical constraints from synthesis through placement through routing. The argument is that capability accumulates through these roles, not within them, and that current approaches plateau at Agent because there is no standard interface for passing design context across tool invocations. The result is that each tool call starts with a smaller model that has already forgotten what the placement step committed to.
The teams building EDA AI infrastructure that will matter in 2028 are not building better generators. They are building context-passing orchestration layers: a structured representation of design state that survives tool handoffs and allows an agent to reason about how a synthesis decision constrains placement options three steps downstream. Labs still benchmarking on RTL generation pass rates are measuring the wrong output. EDA AI startups that have not yet added a context layer to their architecture have 12-18 months before the ones that did pull permanently ahead.