Tag
This paper proposes a lightweight, signal-based framework for efficiently triaging agentic interaction trajectories by computing low-cost indicators that identify informative samples without impacting online agent behavior, achieving an 82% informativeness rate on benchmarks.