VeriTrace Achieves 100% Verilog RTL Generation Accuracy

Current state-of-the-art multi-agent systems for automated Verilog RTL generation typically plateau at around 95% accuracy on standard benchmarks. This limitation stems from an incomplete debugging action space, where agents are restricted in their ability to inspect signals, query time windows, or both. This forces debugging into a pattern-matching approach rather than true hypothesis-driven root-cause analysis. VeriTrace, a new multi-agent system, addresses this by equipping its Inspector agent with a complete debugging action space. This "Agentic Temporal Exploration" allows independent control over signal selection, time-window bounds, and iteration depth, mimicking how human verification engineers explore and debug circuits. By enabling agents to form hypotheses, query waveforms for evidence, and iteratively refine their understanding, VeriTrace has achieved 100% Pass@1 on the VerilogEval-V2 benchmark, marking the first system to reach perfect functional correctness. Even with the same Claude Sonnet 4.0 backbone, VeriTrace significantly outperforms existing baselines, demonstrating the critical role of a comprehensive debugging agency.