A new arXiv paper presents VeriTrace, a multi-agent system for debugging AI-generated Verilog hardware designs. The authors argue that existing systems plateau because agents are not allowed to inspect enough signals, time windows, or iterations during debugging.
VeriTrace gives its Inspector agent independent control over signal selection, time-window bounds, and iteration depth. The paper calls this Agentic Temporal Exploration because the agent can form hypotheses, query waveform evidence, and refine its diagnosis in a way closer to human verification work.
The authors report 100% Pass@1 on VerilogEval-V2, the first perfect functional-correctness result they cite for the benchmark. On a shared Claude Sonnet 4.0 backbone, VeriTrace outperformed the strongest reproduced baseline by 5.1 percentage points.
The result is benchmark-specific, but it shows that agent capability can depend as much on the allowed debugging action space as on the underlying language model.