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

Summary

This arXiv paper (2603.25719) by Abhishek Bhandwaldar, Mihir Choudhury, Ruchir Puri, and Akash Srivastava empirically studies how far general-purpose coding agents, without hardware-specific training, can go in hardware optimization. The authors introduce Agent Factory, a two-stage pipeline for building and orchestrating multiple autonomous optimization agents. Using Claude Code, the system was evaluated on 12 kernels from the HLS-Eval and Rodinia-HLS benchmarks. Scaling from a single agent to 10 agents yielded an average 8.27x speedup, with larger gains on harder benchmarks: streamcluster achieved speedups exceeding 20x. The work demonstrates that coordinated fleets of general-purpose agents can meaningfully optimize high-level synthesis code without domain-specific fine-tuning, offering a practical recipe for hardware design automation.