Regression debugging is becoming harder to scale as semiconductor designs expand to millions of gates and failures multiply across many simulations, test cases, and seeds. The central challenge is no longer only finding one difficult bug, but determining how many distinct problems a regression actually contains, which failures are related, and which simulations carry the information needed to resolve them efficiently.
That pressure is especially visible when bugs require long traces through design hierarchies, long pipelines, large waveforms, and multiple interacting components. In one described case, a bidirectional serialization interface failed after a subordinate triggered reset during a transaction while the manager kept transmitting beyond the reset grace period. Data in the transmitted payload matched the protocol’s start code, causing the subordinate to interpret it as a new frame and leading to a downstream failure. Several conditions had to align before the bug appeared, illustrating why simple question-answering approaches are insufficient for root cause analysis.
The proposed approach treats root cause analysis as a search problem. Instead of exploring every failure independently, the system follows high-probability traces, evaluates different hypotheses, ranks possible explanations, and attaches confidence scores. A secondary pass after conventional deterministic binning is used to group failures that may share a common root cause even when their log signatures differ.
Selectivity is also applied to waveform generation. Rather than rerunning every failing simulation with waveform dumping enabled, the workflow identifies a minimum set of witness simulations that can explain each failure bin. This reduces the amount of expensive rerun activity and limits the data generated to what is needed to debug the regression effectively.
Once the failures have been reduced into meaningful groups and the witness simulations have been collected, the debug packages are handed into autonomous RCA. The system investigates each bin, proposes fixes, verifies them, and uses the verification result together with RCA confidence scoring to decide whether a final regression rerun is warranted. The resulting flow is: regression, intelligent binning, witness selection, waveform rerun, autonomous RCA, fix, verification, and final regression.
The article also emphasizes the cost side of multi-agent systems. Parallel agent exploration can improve performance and accuracy, but the goal is to keep token and compute costs comparable to a traditional single-agent setup. The broader objective is an autonomous debugging workflow that can handle difficult bugs across an entire regression while deciding what to investigate, what to rerun, and where to spend compute.
In that model, the engineer no longer starts triage from scratch after an overnight regression. Instead, the system can resolve some failures autonomously and present the remaining unresolved issues in a structured report for further investigation.