Agentic LLMs for Microarchitecture Research: The Role of the Human Architect

Abstract

Agentic Large Language Models (LLMs) are rapidly becoming an important tool for architecture research, raising a concrete question for the computer architecture community: can an architect simply ask an agent to design a novel mechanism with a better score (e.g., IPC) and expect a meaningful answer? In this work, we study two modes of agentic use on the same microarchitectural domain with the same agentic model: unguided, in which the agent receives a problem statement, a curated seed library, a scoring function, and constraints, and guided, in which a human domain expert additionally reviews iterations and redirects the search when it stalls or drifts. Our evaluation shows that the unguided agent produces a design that performs close to, but below, the state-of-the-art, while the guided agentic design outperforms it. From this contrast, we distill a practical recipe for using today’s agents in microarchitecture research: they are most effective not as autonomous architects but as high-throughput collaborators whose output improves substantially under a human architect’s steering.