Elastic Horizon tracks the 90th percentile of successful trajectory lengths to stop agents padding turns, saving 25% of tokens

arXiv (2609.07247)·medium signalXBlueskyLinkedInCopy link

arXiv 2609.07247 proposes an 'effective interaction frontier' past which extra agent interactions yield diminishing returns while cost keeps growing linearly, then builds a closed-loop controller that locates that frontier from the 90th percentile of successful trajectory lengths. On AppWorld and BFCL it gets the best success rates on both 7B and 14B backbones while saving up to 25% of per-step trajectory tokens. It reaches matched success with strictly fewer cumulative tokens than every fixed-horizon baseline, which removes the manual horizon tuning step.