WebSwarm: Recursive Multi-Agent System Beats Single-Agent Web Search
New framework uses recursive delegation for deep-and-wide web search tasks.
WebSwarm addresses the limitations of single ReAct-style agents in handling deep-and-wide web search. A single long trajectory and limited context make it hard to simultaneously cover depth and breadth. Existing multi-agent systems improve coverage through parallel execution but lack recursive depth and evidence-grounded expansion.
WebSwarm introduces a progressive recursive delegation framework. It probes task-relevant information organization on the web, then instantiates agentic search nodes with local objectives and search modes. Nodes can delegate child nodes or solve their own objective, returning evidence upward. Experiments on multiple benchmarks show WebSwarm outperforms baselines, with ablation studies confirming its effectiveness.
- WebSwarm uses recursive delegation to dynamically instantiate agentic search nodes that can spawn child nodes for deeper exploration.
- It outperforms single-agent and multi-agent baselines on BrowseComp-Plus, WideSearch, DeepWideSearch, and GISA benchmarks.
- The framework reuses process-level experience across homogeneous sibling nodes to improve efficiency.
Why It Matters
By combining recursive depth with parallel coverage, WebSwarm enables more thorough, research-grade web search from AI agents.