Researchers explore methods for detecting and attributing failure in LLM agents, finding that internal state monitoring outperforms traditional methods. The study suggests that current benchmarking practices may be misleading, and that gains in performance may be due to regularities in the data rather than actual improvements in agent performance.