AutoCircuit: Agentic Pareto Front Exploration for Analog Circuit Designs

Analog circuit design poses a high-dimensional search problem in which circuit parameters must be jointly optimized with respect to competing performance targets such as gain and power. A range of traditional AI methods have shown promising results on this task, including Bayesian optimization (BO), evolutionary algorithms, and reinforcement learning. More recently, large language models (LLMs) have been proposed as autonomous design agents. However, prior work typically studies these techniques in isolation rather than composing them and often focuses on scalarized performance metrics, which can lead to limited design diversity. To address this, we present \textbf{AutoCircuit}, a system with an LLM orchestrator that leverages a set of ML tools to explore the Pareto front in a multi-objective space within a fixed wall-clock budget. We evaluate our system on a representative analog benchmark, a two-stage op-amp circuit implemented in the SKY130 open PDK, and show that LLM orchestration with effective tool use improves log-hypervolume by up to $2.4\times$ over the LLM-only baseline. We further show that the best result comes from a carefully chosen subset of tools (NSGA-II~+~$g_m/I_d$) rather than the full set of tools.