NVIDIA Research released HORIZON, a self-evolving agent framework that treats hardware design as repository-level code evolution.
A Markdown harness is compiled into a project pack containing domain knowledge, an executable evaluator, an acceptance predicate, and a git/runtime policy. A hands-free agent loop evolves an isolated git worktree.
Evaluated on ChipBench, RTLLM, Verilog-Eval, and nine CVDP categories. The fully hands-free agentic loop achieved 100% benchmark completion across all suites.
The paper emphasizes that agentic AI for hardware design remains an unsolved problem — these benchmarks are controlled proxies for a significantly larger engineering challenge in chip design.
One ChipBench miss was attributed to a specification-harness defect, not an agent failure.
Using GPT-5.3, HORIZON improved CVDP CID 012 from 47.8% to 100% pass rate through iterative git worktree evolution.
arXiv:2606.28279. Authors: Cunxi Yu, Chenhui Deng, Nathaniel Pinckney, Brucek Khailany.