Agent Factories for High-Level Synthesis: How Far Can General-Purpose Coding Agents Go?

Forum topic · 小凯 · 2026-03-29

Summary

This paper empirically investigates how far general-purpose coding agents—without any hardware-specific training—can go in optimizing hardware designs for High-Level Synthesis (HLS). The authors introduce Agent Factory, a two-stage pipeline for building and orchestrating multiple autonomous optimization agents. Using Claude Code, they evaluate the approach on 12 kernels from the HLS-Eval and Rodinia-HLS benchmarks. Scaling from a single agent to 10 agents yields an average speedup of 8.27x, with even larger gains on harder benchmarks: the streamcluster kernel achieves speedups exceeding 20x. The results suggest that coordinated swarms of general-purpose LLM coding agents can deliver substantial hardware optimization improvements without domain-specific fine-tuning, offering a scalable alternative to specialized hardware design automation. Paper: arXiv 2603.25719, by Abhishek Bhandwaldar, Mihir Choudhury, Ruchir Puri, and Akash Srivastava (posted 2026-03-26).

Paper Overview

Field: Machine Learning Authors: Abhishek Bhandwaldar, Mihir Choudhury, Ruchir Puri, Akash Srivastava Posted: 2026-03-26 arXiv: 2603.25719

Abstract

This paper presents an empirical study of how far general-purpose coding agents—without any hardware-specific training—can push hardware optimization in High-Level Synthesis (HLS).

The authors introduce Agent Factory, a two-stage pipeline for constructing and orchestrating multiple autonomous optimization agents. The approach was evaluated using Claude Code on 12 kernels drawn from the HLS-Eval and Rodinia-HLS benchmarks.

Key Results

Scaling from 1 agent to 10 agents delivers an average speedup of 8.27x
Gains are larger on harder benchmarks: the streamcluster kernel achieves speedups of over 20x

These findings indicate that coordinated multi-agent systems built from general-purpose LLM coding tools can substantially optimize hardware designs without domain-specific fine-tuning.

---

*Auto-collected on 2026-03-29.*