HORIZON: Treating Hardware Design as Code Evolution

February 18, 2026

I came across a paper recently that, although it's fairly technical, I think is worth sharing because the core idea is genuinely interesting.

The paper (arXiv:2606.28279, published June 26, 2026) is called "Agentic Hardware Design as Repository-Level Code Evolution" and introduces HORIZON, an AI agent framework for hardware design built by NVIDIA Research (Cunxi Yu, Chenhui Deng, Nathaniel Pinckney, and Brucek Khailany).

The core idea

Instead of treating RTL (Verilog) code generation as a one-shot task — prompt in, code out — HORIZON treats hardware design as code evolution at the repository level, the same way modern software development works.

In practice, this means an AI agent edits an isolated git worktree, runs an automated evaluator (compilation, simulation, and coverage), and either commits or discards each change depending on whether it passes the tests.

How the workflow works

Roughly:

  1. The user writes a Markdown harness describing the objective.
  2. A bootstrap agent converts it into a Project Pack: domain knowledge, an executable evaluator, acceptance criteria, and Git policies.
  3. A fully autonomous agent loop (no human intervention) edits, evaluates, commits, or discards successive versions.

What I find interesting is the role Git plays here: it's not just version control, it becomes the agent's memory and traceability layer, through diffs, commits, and logs.

Results

The framework was evaluated on ChipBench, RTLLM-2.0, Verilog-Eval, and the nine CVDP categories (a more demanding RTL design and verification benchmark), achieving a 100% pass rate across all evaluation suites using GPT-5.3 as the base model.

A few details that stood out to me:

  • Simpler tasks converge in just 2-5 iterations.
  • The hardest CVDP tasks (like code completion or generating verification checkers) require significantly more iterations — up to 82.
  • Token consumption is heavily concentrated in the hardest categories.
  • About 91% of tokens come from cache thanks to session reuse, which significantly cuts computational cost.

In summary

This is a technical paper that proposes managing hardware design as if it were software evolution inside a Git repository, driven by fully autonomous AI agents. The authors themselves are careful to note that this doesn't mean agentic AI for hardware design is solved: these benchmarks are controlled proxies for a much broader engineering problem.