Web Search Agents

Web search agents study how language models can plan searches, navigate webpages, gather evidence, and produce answers or complete tasks that require current information. Common approaches include query refinement, structured retrieval, persistent workspaces and context management, parallel or hierarchical agents, and training with supervised trajectories, reinforcement learning, or self-play. Benchmarks such as BrowseComp, WideSearch, and live or long-horizon evaluations test whether agents can find broad, timely evidence and use it accurately. A central challenge is that finishing a search or producing a plausible answer does not ensure complete, well-grounded results, especially when evidence is unreliable or benchmarks reward facts a model already knows.