Submitted by Hengjie Zhu 35 Trial by Trace: Benchmarking AI Agent Harnesses for Reliability and Failure Attribution TraceLite AI 28 6