Evaluating Agentic AI in the Wild: Failure Modes, Drift Patterns, and a Production Evaluation Framework
arXiv:2605.01604v1 Announce Type: new
Abstract: Existing evaluation frameworks for large language models — including HELM, MT-Bench, AgentBench, and BIG-bench — are designed for controlled, single-session, lab-scale settings. They do not address the…