NVIDIA HORIZON Agent Automates RTL Hardware Design With 100% Benchmark Completion

NVIDIA Research has released HORIZON, an autonomous AI agent framework that automates Register-Transfer Level (RTL) hardware design by treating each problem as repository-level code evolution, according to TechGig. The system achieved a 100% completion rate across all evaluated RTL benchmark suites, including ChipBench, RTLLM-2.0, Verilog-Eval, and multiple CVDP verification categories.

The team is explicit that “agentic hardware design is not yet a fully solved problem,” but the benchmark results suggest agents can now handle a meaningful range of daily RTL work.

How HORIZON Works

HORIZON defines each design problem as a version-controlled repository, moving beyond single-turn code generation. The primary input is a structured Markdown harness specifying the goal, domain-knowledge directions, an evaluator specification, and an acceptance predicate, TechGig reports.

A bootstrap agent compiles this harness into a project pack that guides the agent’s policy, executable evaluator, acceptance predicate, version-control policy, and domain skills. The agent loop then operates without human intervention: planning targets, editing the worktree, invoking tools, and running the evaluator.

Git serves as the core substrate. The system commits new versions only when acceptance gates pass, and treats accepted and rejected attempts as positive and negative repair examples. Repository history becomes an experience buffer, giving the agent context on what worked and what failed without retraining.

Benchmark Results and Token Economics

Using GPT-5.3 as its backbone, HORIZON demonstrated 100% pass rates across RTL code completion, natural-language-to-RTL conversion, code modification, module reuse, linting, quality-of-results improvement, verification generation, and functional debugging, according to TechGig.

The convergence difficulty varied significantly by category. Checker generation started at 3.8% and climbed to 100% within 19 iterations. The system’s ability to iterate autonomously through failure cases is core to the architecture.

On cost: approximately 91% of all tokens consumed were cached input, significantly reducing API costs. The research team identifies token efficiency, not final pass rate, as the primary area for future improvement.

Agents in Specialized Engineering

HORIZON represents a pattern that extends beyond hardware design. The framework treats complex engineering problems as repository evolution tasks where agents iterate, fail, learn from history, and converge. This is the same loop that coding agents like Claude Code and Codex CLI already run for software, now applied to hardware description languages like Verilog and SystemVerilog.

For chip design teams, the practical implication is acceleration. RTL design involves repetitive iteration cycles of writing, simulating, debugging, and rewriting. An agent that handles these cycles autonomously, with cost controlled through aggressive caching, changes the calculus on which design tasks require human engineers and which can be automated.

The research team emphasizes that HORIZON addresses daily RTL work, not novel architecture design. The 100% benchmark completion demonstrates competence on established patterns, not creativity on unsolved problems.