Agentic applications based on large language models increasingly rely on multi-step interaction loops involving planning, action execution, and environment feedback. Improving them post-deployment remains challenging because agent trajectories are voluminous and non-deterministic, making human or auxiliary LLM review slow and cost-prohibitive. (arXiv:2604.00356)

The paper proposes computing cheap signals from live interactions attached as structured attributes for trajectory triage to identify informative interactions. Signals are organized into a taxonomy spanning interaction (misalignment, stagnation, disengagement, satisfaction), execution (failure, loop), and environment (exhaustion), designed for computation without model calls. (arXiv:2604.00356)

A controlled annotation study on τ-bench showed signal-based sampling achieves an 82% informativeness rate compared to 74% for heuristic filtering and 54% for random sampling, with a 1.52x efficiency gain per informative trajectory. The advantage is robust across reward strata and task domains. (arXiv:2604.00356)