New Survey Maps the Landscape of AI Agents in Chip Design Workflows

Electronic design automation has long been a multi-layered, handoff-intensive process where engineers juggle transfers between people, design files, and specialized software. A new survey from arXiv now reveals how language model-based agents are increasingly stepping into these transfer roles—yet the field remains scattered and lacks a coherent framework.

A New Lens for Understanding Agentic EDA#

Researchers have developed a taxonomy centered on what moves between stages: the primary input and output for each agent system. Rather than treating all agentic systems alike, they classify them into three categories. Intra-stage systems operate within a single phase of the design flow. Inter-stage systems convert design artifacts from one phase into formats the next phase can consume. Extra-stage systems take high-level human intent and translate it into usable EDA artifacts.

This input-output perspective reveals a fundamental structural pattern: who initiates work (the source of the primary input) and who ultimately benefits from it (the downstream consumer) determine how each agent fits into the broader design ecosystem.

Scope and Findings#

The survey examines 115 representative agentic systems across multiple dimensions—research trends, evaluation benchmarks, and the technical mechanisms underlying these agents. By organizing the literature this way, the authors expose gaps where research clusters densely and where significant challenges remain unaddressed.

Why This Matters#

As LLM-based agents become active participants in chip design, fragmentation becomes costly. Engineers and researchers lack a unified vocabulary for understanding where agents add value, how they should hand off to one another, and what standards they need to follow. The new taxonomy and roadmap suggest potential protocols that could streamline integration and reduce friction as automation deepens across EDA workflows.

The full analysis, including an agentic EDA protocol roadmap and discussion of open research problems, is available on arXiv.