WebClipper: Efficient Evolution of Web Agents with Graph-based Trajectory Pruning

Deep Research systems leveraging web agents face challenges in search efficiency due to long tool-call trajectories, cyclic reasoning, and unproductive explorations. WebClipper is a novel framework that addresses this by compressing web agent trajectories through graph-based pruning. It models the agent's search as a state graph and optimizes trajectories into minimal Directed Acyclic Graphs, preserving essential reasoning while eliminating redundancy. Continuous training on these refined trajectories allows agents to evolve more efficient search patterns, reducing tool-call rounds by approximately 20% while improving accuracy. This work provides practical insights into balancing effectiveness and efficiency in web agent design.