Agentic AI Software Debugging in OpenBMC: Extracting Ground Truth from Binaries to Eliminate Hallucinations

With the rapid development of AI, its applications in the embedded software field are increasingly widespread. However, debugging OpenBMC in production environments remains notoriously difficult. Runtime logs are overwhelmingly long, making routine inspections highly time-consuming, and traditional troubleshooting relies heavily on manual analysis to reconstruct causal relationships. Even with current Agentic assistance, AI tools typically rely on superficial text-matching or vector embeddings to retrieve context. Because these methods lack deep semantic understanding of the codebase, directly feeding raw logs to AI agents rapidly exhausts context windows and produces factually incorrect diagnoses.
This talk introduces a non-intrusive, zero-hallucination Agentic AI debugging workflow for OpenBMC. Instead of guessing logic, the Agent extracts deterministic ground truth from compiled artifacts. By leveraging binary analysis tools to pre-analyze symbol-rich binaries, we extract accurate call chains and static log strings. When a crash occurs, the Agent automatically captures runtime logs, reverse-matches isolated log lines to their precise source context, and traverses accurate call graphs to reconstruct the execution flow. By feeding the LLM this highly condensed, deterministic evidence chain rather than massive log dumps and raw source files, the Agent focuses solely on logical deduction.
This approach requires zero modifications to the OpenBMC build process, ensures non-intrusiveness, significantly reduces token consumption, and enables end-to-end automated, highly accurate root cause analysis (RCA).