SearchHorizonBench Measures Search Value in AI Agents
October 25, 2026
SearchHorizonBench evaluates an agent's ability to estimate if additional search execution is worth its cost. The benchmark uses 45,632 executions to show that historical calibration can reduce Brier error by 16.2% in predicting search success.
HOW THIS AFFECTS YOU
●
builderYou can use these findings to optimize agentic workflows by predicting when a search step is likely to yield high utility.
●
researcherThis provides a framework to separate generation noise from true search-gain estimation.