What VAKRA Reveals About Why Agents Actually Fail

VAKRA, a new benchmark from IBM Research, reveals that AI agents fail in predictable, structural ways by mapping fracture points between reasoning, tool selection, and execution. It decomposes agent failure into six specific categories, moving beyond traditional binary task completion evaluations to uncover common weaknesses.