The Big Picture
Recursive delegation that spawns specialized search nodes and uses lightweight probing of web structure lets a team of search agents achieve more accurate and more thorough web information retrieval than single-agent or fixed-collaboration systems.
ON THIS PAGE
The Evidence
A search system that dynamically creates and delegates specialized sub-agents as evidence appears handles both deep multi-step reasoning and broad coverage better than one-size-fits-all approaches. Each sub-agent gets a local objective plus a search mode (solve locally or spawn children), and the system decides further expansion based on returned evidence. web-structure probing (to see if information is concentrated or dispersed) and reusing experience from similar sibling nodes reduce redundant work and guide better decompositions. Across multiple benchmarks, this approach yields consistent gains, particularly on the hardest tasks that need nested depth and wide coverage.sub-agent delegation
Data Highlights
1Evaluated on 4 challenging benchmarks: BrowseComp-Plus, WideSearch, DeepWideSearch, and GISA.
2For BrowseComp-Plus experiments, 200 instances were sampled for resource-efficient evaluation.
3Consistent outperformance of single-agent and prior multi-agent baselines across all benchmarks, with especially clear gains on the most difficult examples (deep-and-wide tasks).
What This Means
Engineers building autonomous web-search agents or assistant systems: use recursive delegation and autonomous web-search agents web-structure probing to avoid overloading single agent trajectories. Technical leads evaluating search pipelines: this shows a practical way to trade coordination complexity for better coverage and fewer dead-end searches. Researchers studying agent collaboration: the paper gives a concrete architecture and ablations to test recursive delegation and experience reuse.
Not sure where to start?Get personalized recommendations
Key Figures
Fig 1: Figure 1: Illustration of representative multi-agent orchestration paradigms and WebSwarm.
Fig 2: Figure 2: Overview of WebSwarm with a running example from the DeepWideSearch benchmark.
Fig 3: Figure 3: Performance and web tool usage of GLM-4.5 across difficulty levels on BrowseComp-Plus and WideSearch-EN.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
Reported results focus on benchmark settings and English subsets, with one dataset sampled for efficiency, so real-world gains may vary on larger or multilingual corpora. The method increases orchestration complexity and may raise run-time and tooling costs compared with single-agent setups. Performance depends on the underlying language model and the quality of web probing; robustness under adversarial or highly noisy web structure was not fully explored.
Methodology & More
WebSwarm treats a web search task as a growing tree of small search nodes rather than as a single monolithic agent run. A root agent starts with the user query and either solves it directly or spawns child nodes; each child receives a narrow local objective and a ‘‘search mode’’ that determines whether it should perform iterative local search or further delegate. Nodes return evidence upward, and parent nodes decide to expand, revise, or aggregate results. Key control ideas are (1) grounding expansion in lightweight probes of web information structure—so the system splits tasks along dimensions that match how information is organized on the web—and (2) reusing short-term experience from earlier sibling nodes to guide later, similar subtasks and avoid repeating mistakes. multi-agent designs Evaluation used four established benchmarks covering deep factual queries, wide structured collection, nested deep-and-wide tasks, and general information seeking. WebSwarm consistently beat both single-agent baselines and prior multi-agent designs by dynamically changing collaboration patterns (parallel split, sequential verify, or multi-path sampling) per node rather than committing to one global strategy. Practical implications: when a search problem reveals new entities or constraints during the process, a progressive, evidence-driven delegation strategy reduces wasted retrieval and improves both completeness and correctness—especially for the hardest tasks that require multi-hop reasoning across many candidate pages.
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Multiple authors with moderate-to-high h-indices (several >15 and a few ~28–29) and a university affiliation (University of Montreal) — established researchers though not top industry labs or top-venue listing.