WebSwarm is a progressive recursive delegation framework designed to address the limitations of single ReAct-style agents in handling complex, deep-and-wide web search tasks. The system dynamically instantiates agentic search nodes that couple local objectives with specific search modes, allowing them to either solve tasks directly or delegate to child nodes.
- WebSwarm jointly constructs task decomposition, recursive expansion, and agent collaboration during inference.
- Nodes return evidence and results upward, enabling parent nodes to expand, revise, or aggregate the search process.
- The framework probes how task-relevant information is organized on the web to ground node expansion and reuses experience across homogeneous sibling nodes.
- Experiments show WebSwarm consistently outperforms single-agent and multi-agent baselines on BrowseComp-Plus, WideSearch, DeepWideSearch, and GISA benchmarks.
This approach enables more effective handling of depth and coverage simultaneously by leveraging recursive delegation and evidence-grounded expansion.