Submitted by Hengjie Zhu 34 Trial by Trace: Benchmarking AI Agent Harnesses for Reliability and Failure Attribution TraceLite AI 25 6