Agentic Test-Time Scaling for WebAgents

ABSTRACT

Current WebAgents struggle with long-horizon tasks and complex navigation. We propose an agentic scaling framework that increases compute at test-time through iterative trajectory pruning and reward-guided search. By treating the agent as a search process over the browser's state space, we show that performance scales predictably with the number of explored paths. Our method significantly outperforms larger, more expensive base models by using a smaller model with an optimized search-and-verify loop.