Agentic Hardware Design as Repository-Level Code Evolution

A team of researchers has introduced HORIZON, a self-evolving agent framework that treats hardware design as repository-level code evolution, according to a paper submitted to arXiv on 26 Jun 2026 [1][2]. The framework compiles a Markdown harness into a project pack containing domain knowledge, an executable evaluator, an acceptance predicate, and a git/runtime policy. A hands-free agent loop then evolves an isolated git worktree, using repository operations for state management, tracing, and replay [1][2]. This extends prior work on repository-scale self-evolution from electronic design automation software systems to hardware-design artifacts themselves [2]. The researchers evaluated HORIZON on ChipBench, RTLLM, Verilog-Eval, and nine CVDP categories, achieving 100% benchmark completion across all suites with a fully hands-free agentic loop [1][2]. The paper does not claim that agentic AI for hardware design is solved, noting that these benchmarks are controlled proxies for a much broader engineering problem in chip design [2]. Repository-level code evolution, the paradigm HORIZON employs, draws on decades of open-source collaboration practices. Linux, first released on 17 September 1991 by Linus Torvalds, is one of the most prominent examples of free and open-source software collaboration and is considered by many to be the largest open source project [3]. The Linux kernel has been ported to more platforms than any other operating system and is used on devices ranging from personal computers to all of the world's 500 fastest supercomputers [3]. HORIZON's approach of treating hardware design artifacts through repository operations mirrors the distributed development workflows that have scaled open-source operating systems. Ubuntu, a Linux distribution based on Debian, is published on a six-month release cycle with long-term support versions issued every two years, demonstrating how structured repository management can sustain large-scale software projects [5]. The paper's submission to arXiv places it within a broader ecosystem of community-driven research dissemination. arXiv itself operates arXivLabs, a framework that allows collaborators to develop and share new features directly on the website, with both individuals and organizations embracing values of openness, community, excellence, and user data privacy [11]. The framework complements other collaborative projects and demonstrates how results are achieved when the community works together [11]. The researchers acknowledge limitations in the current study and highlight open research challenges in a dedicated discussion section [2]. The work represents a step toward automated hardware design workflows but remains bounded by the gap between benchmark performance and the full complexity of industrial chip design [2].