The Tool Illusion: Rethinking Tool Use in Web Agents

arXiv cs.CL · 6 hours ago

Researchers conducted a large-scale empirical study examining tool use in web agents, testing multiple tool sources, models, and evaluation benchmarks to clarify whether tools consistently improve performance. The study evaluated tools across diverse experimental settings and frameworks to establish comparable baselines that prior work had lacked. The findings provide clearer evidence about tool design principles and potential drawbacks, establishing a more reliable foundation for developing web agents that use tools.