NVIDIA HORIZON: A Hands-Free Agent that Evolves Git Worktrees and Hits 100% RTL Benchmark Completion
NVIDIA Research introduces HORIZON, a hands-free agentic framework that treats hardware design as repository-level code evolution using isolated Git worktrees. The system achieves a 100% pass rate across all evaluated RTL benchmark suites (ChipBench, RTLLM-2.0, Verilog-Eval, and CVDP) by leveraging iterative self-correction. HORIZON utilizes a structured Markdown harness to define goals, evaluators, and acceptance predicates, compiling them into a self-contained agent loop that commits only upon
Analysis
TL;DR
- NVIDIA Research introduces HORIZON, a hands-free agentic framework that treats hardware design as repository-level code evolution using isolated Git worktrees.
- The system achieves a 100% pass rate across all evaluated RTL benchmark suites (ChipBench, RTLLM-2.0, Verilog-Eval, and CVDP) by leveraging iterative self-correction.
- HORIZON utilizes a structured Markdown harness to define goals, evaluators, and acceptance predicates, compiling them into a self-contained agent loop that commits only upon passing executable gates.
- The framework demonstrates significant variance in convergence speed, ranging from 2 iterations for simple tasks to 82 iterations for complex code completion, highlighting the computational cost of iterative refinement.
Why It Matters
This development marks a critical shift from single-turn code generation to iterative, autonomous engineering workflows for hardware description languages (HDLs). By proving that agentic systems can achieve 100% success rates on complex RTL benchmarks through continuous self-evaluation, it validates the feasibility of fully automated hardware design pipelines. This reduces the manual burden on engineers and establishes a new standard for reliability in AI-assisted chip design.
Technical Details
- Repository-Level Evolution: HORIZON manages designs as version-controlled Git repositories rather than static files. Each iteration involves planning, editing the worktree, invoking tools (compilation, simulation, coverage), and committing only if an executable acceptance predicate passes.
- Structured Harness & Project Pack: Input is defined via a Markdown harness containing goals, domain knowledge, evaluator specs, and acceptance predicates. A bootstrap agent compiles this into a project pack $p = (\pi_{agent}, E_p, A_p, \Gamma_p, \Omega_p)$, defining the agent policy, evaluator, acceptance predicate, version-control policy, and domain skills.
- Git as Experience Buffer: The system uses native Git commands to trace trajectories. Successful commits serve as positive repair examples, while rejected attempts are logged as negative examples. The repository history acts as the experience buffer, eliminating the need for separate datastores.
- Fixed Backbone & Cost Optimization: Experiments use a fixed GPT-5.3 backbone without reinforcement learning updates. Token costs are minimized by maintaining persistent model sessions and leveraging prompt caching for stable sources, billing only for diffs and evaluator outputs.
- Benchmark Performance: Evaluated on ChipBench, RTLLM-2.0, Verilog-Eval, and CVDP (783 problems). While first-iteration pass rates vary (e.g., 3.2% for RTL code completion vs. 86.2% for Verilog-Eval), all suites reach 100% final pass rates after sufficient iterations.
Industry Insight
- Adoption of Iterative Agentic Workflows: Hardware design teams should anticipate a move toward multi-step, self-correcting AI agents rather than one-shot generation tools. Integrating Git-based version control into AI workflows becomes essential for tracking and validating these iterative improvements.
- Resource Planning for Long-Tail Tasks: The wide variance in convergence iterations (from 2 to 82) indicates that certain design tasks, like code completion or checker generation, will incur significantly higher computational costs. Organizations must budget for extended inference times and token usage for complex verification tasks.
- Standardization of Acceptance Predicates: The success of HORIZON relies heavily on precise, executable acceptance criteria. Developing robust, automated verification suites (simulators, linters, coverage tools) is now a prerequisite for effectively deploying agentic AI in hardware engineering.
Disclaimer: The above content is generated by AI and is for reference only.