General-Purpose Coding Agents Optimize Hardware Designs via Agent Factory Pipeline

A recent investigation published on arXiv (2603.25719) examines the capability of general-purpose coding agents to enhance hardware designs based solely on high-level algorithmic specifications, without requiring hardware-specific training. The researchers present an "agent factory," which consists of a two-phase process that develops and manages several autonomous optimization agents. In the initial phase, the design is broken down into sub-kernels, with each being optimized independently through pragma and code-level transformations. An Integer Linear Program (ILP) is then created to compile the most promising configurations while adhering to area constraints. The second phase involves deploying N expert agents to analyze the top ILP solutions, focusing on cross-function optimizations like pragma recombination, loop fusion, and memory restructuring. Evaluations conducted on 12 kernels from HLS-Eval and Rodinia-HLS, utilizing Claude Code (Opus 4.5/4.6) alongside AMD Vitis HLS, demonstrated performance enhancements as the number of agents increased from 1 to 10. This research indicates that general-purpose coding agents can effectively optimize hardware designs, potentially minimizing the necessity for specialized knowledge in high-level synthesis.