Hirain argued AI agents fail in vehicle development for want of machine-readable rules

Beijing Hirain Technologies published a technical account on October 14 arguing that general-purpose AI agents fail inside automotive development for reasons that have little to do with model capability. It opened on a gap it said was visible at carmakers and tier one suppliers alike: more AI tools in use, and development cycles compressing out of proportion to that use.

The company set out three obstacles. The first was the rule harness that constrains what an agent may do, which it said grows with vehicle complexity until maintaining it costs more labor than it saves, a dozen or so carriers written up rule by rule easily passing a thousand rules. In the context of functional safety under ISO 26262 and the UN R155 and UN R156 cybersecurity regulations, Hirain wrote, an audit asks which clause applied and how the judgment was reached, not whether the model understood correctly on the day.

The second was open-loop output. Handing a functional specification to a general-purpose agent returned neatly formatted test cases in seconds, it said, with a strikingly high proportion sent back at review, no return path for the review comments, and the same error recurring as a result. Its own domain rule library drew its criteria from tens of thousands of manually written test cases.

The third was retrieval. Hirain argued that retrieval-augmented generation assumes static knowledge, while automotive electronics is a field where requirements change, the communication matrix is reissued and regulation iterates. Its sharpest claim followed: running an agent on a knowledge base with no version history amounted to making a safety argument on an uncontrolled configuration item, citing ISO 26262 on traceable evidence and ASPICE SUP.8 and SUP.10 on configuration control. The piece closed by pointing to a seminar on October 15.