An open-source AI workflow can now refactor general software into code that hardware compilers can actually use — faster and cheaper than prior methods.
High-Level Synthesis (HLS) is a shortcut for chip designers: write something close to software, get hardware out the other side. The catch is that real-world code rarely survives the translation without heavy rewriting. AgRefactor, introduced in a new paper, uses a multi-agent LLM system to automate that rewriting. It adds a self-evolving memory layer so the system gets smarter across tasks, and it offloads repetitive changes to deterministic tools rather than burning LLM calls on everything. On 9 of 11 benchmarks — programs five to ten times longer than anything prior work attempted — it matched or beat both the leading automated refactoring tool and a strong LLM-only baseline.
The performance numbers are the headline: a 6.51x geometric mean speedup over the state-of-the-art pragma tuning tool, and a 1.20x improvement over optimized open-source designs, at less than 20 percent extra resource cost. That matters because HLS bottlenecks are real — chip design cycles are long, and anything that compresses the software-to-silicon gap has direct value for teams building custom accelerators.
The self-improving memory angle is the part worth watching. Most LLM coding pipelines are stateless; AgRefactor accumulates strategic knowledge across runs, which is either a genuine efficiency gain or a vector for compounding errors — the paper claims the former, but production deployments will tell the real story.