Some automotive bugs are invisible to every tool that reads the code. A signal that two ECUs interpret differently, or a value mapping that got swapped somewhere along the way, is perfectly valid code by any syntactic measure. The compiler passes it, the linter passes it, the static analyzer passes it, and so does the review. It turns up much later, on real hardware, when someone has to work backwards from a symptom.

AUMOVIO's problem finder looks for these issues earlier. It runs on Amazon Bedrock AgentCore and is built with the Strands Agents SDK: an orchestrator agent, plus analysis agents that work in parallel, each checking the codebase against a different problem template. The templates are based on issues AUMOVIO has already hit, so the system is looking for this codebase's known failure modes rather than anyone's idea of clean code.

How well that works depends on the template. On one, experts confirmed all six findings. On another, three out of eight. A scan of a safety-relevant subsystem that had already been through review turned up more than thirty high-priority issues. Findings come out shaped like Jira tickets, and a scan over changed code takes minutes, which is short enough to sit in a pull request.