NVIDIA HORIZON Agent Achieves 100% RTL Benchmark Completion in Hands-Free Hardware Design

NVIDIA Research has unveiled HORIZON, an innovative hands-free AI agent framework designed to automate and evolve hardware designs. This new system treats hardware design as a process of repository-level code evolution, leveraging Git worktrees to autonomously iterate and refine Register-Transfer Level (RTL) instantiations. Notably, HORIZON has achieved a 100% completion rate across all evaluated RTL benchmark suites, marking a significant advancement in automated hardware development.

The HORIZON framework operates by hosting each hardware design problem as a version-controlled repository. The core input is a structured Markdown harness that outlines the design goal, domain-knowledge directions, an evaluator specification, and an acceptance predicate. A bootstrap agent compiles this harness into a project pack, which then guides a self-contained agent loop. This loop continuously evolves an isolated Git worktree, committing new versions only when an executable acceptance gate passes.

Key facts

Feature	Description
Framework Name	NVIDIA HORIZON
Core Function	Hands-free AI agent for hardware design evolution
Key Technology	Git worktrees for version control and iteration
Performance	100% completion across all evaluated RTL benchmark suites
Backbone Model	GPT-5.3

Autonomous Design Evolution

Unlike traditional single-turn code generation methods, HORIZON’s approach emphasizes continuous evolution. The agent’s loop involves planning a target, editing the worktree, invoking design tools, and running an evaluator. The acceptance predicate then determines whether to commit the new version or log a failure. Git serves as the fundamental substrate, allowing for detailed tracing of proposed state changes through diffs, defining accepted checkpoints via commits, and attaching evaluator evidence through notes. This full trajectory can be recovered from the log, making the process transparent and auditable.

The framework’s persistent model session across iterations helps maintain cost efficiency. While the agent backbone (GPT-5.3) remains fixed, newly billed tokens are primarily driven by the current diff and the latest evaluator output, with about 91% of all tokens being cached input. This highlights a focus on token efficiency as a key metric for improvement.

Impact on Hardware Development

The research team notes that while HORIZON achieves high completion rates, agentic hardware design is not yet fully solved. The complexity of real hardware correctness, dependent on cycle-level behavior, reset conventions, bit widths, and simulator feedback, goes beyond plausible Verilog. However, HORIZON’s success in automating generation, completion, and repair across entire suites demonstrates a powerful new paradigm. It extends the lineage of repository-scale self-evolution, moving from software to hardware artifacts.

The evaluation included major benchmark suites such as ChipBench, RTLLM-2.0, and Verilog-Eval, alongside nine CVDP code- and verification-generation categories. While some categories, like checker generation, started with low first-iteration pass rates (3.8%), HORIZON consistently iterated to reach 100% completion, showcasing its robust repair and debugging capabilities. This iterative approach, driven by commercial-EDA simulation, allows the agent to refine designs until specifications are met.

Relevance for Indian Tech and Startups

For the Indian technology and startup ecosystem, the development of NVIDIA HORIZON presents significant opportunities. As India continues to strengthen its semiconductor and hardware design capabilities, tools that automate complex RTL tasks can accelerate product development cycles and reduce human error. Startups focusing on AI-driven design automation or chip manufacturing could leverage such frameworks to enhance efficiency and competitiveness. The ability to achieve 100% completion in hands-free mode means that Indian engineers could potentially focus on higher-level architectural challenges, leaving iterative design refinements to AI agents.

Source: MarkTechPost (https://www.marktechpost.com/2026/07/04/nvidia-horizon-a-hands-free-agent-that-evolves-git-worktrees-and-hits-100-rtl-benchmark-completion/)