A paper being presented at DAC 2026 (opening today in Long Beach) benchmarks three agent architectures and four foundation models against a PicoRV32 RTL-to-GDS flow using commercial EDA tools. The answer to the title question "Can AI agents really complete RTL-to-GDS?" is: not reliably, and the bottleneck is not model capability.
The benchmark produces two findings that matter for anyone building agentic EDA. First, agents achieving similar design completion scores differ by up to 141x in Token ROI, the ratio of design quality to inference cost. "Capable of completing the task" and "cost-efficient at completing the task" are independent axes, and the gap between them is not small. Second, the dominant failure mode is low-level tool-interface mismatch: Tcl commands that behave differently depending on tool version or execution context. Domain-specific skills improved individual subtask execution but did not prevent failure at the hand-offs between stages. Long-horizon EDA flows fail at the seams, not at the hard parts.
The paper's prescription is structured tool interfaces, persistent design context, and process-level evaluation. This is an infrastructure problem, not a model problem. A team building agentic EDA on well-typed, versioned, programmatic tool interfaces will produce far better Token ROI than a team building on brittle Tcl wrappers, regardless of which foundation model it uses. The EDA vendor that ships a structured API layer first, one that is version-stable, observable, and callable from an agent context, owns the agentic design infrastructure layer. The model providers are not the constraint here. The tool interfaces are.