Unclear advantages of multi-agent versus single-agent systems for deep research

Determine which evaluation dimensions favor multi-agent deep research systems over single-agent systems, identify the fundamental bottlenecks in multi-agent approaches, and establish whether multi-agent systems provide consistent gains relative to single-agent systems on dynamic, multi-faceted deep research tasks.

Background

The paper surveys architectures for deep research agents, contrasting single-agent systems (one model making all tool-use decisions) with multi-agent systems that coordinate specialized roles such as planner, researcher, and writer. While multiple proprietary and open-source systems exist, standardized comparisons have been difficult due to varying benchmarks and evaluation methodologies.

Using LiveResearchBench and DeepEval, the authors evaluate 17 systems spanning single-agent web search, single-agent deep research, and multi-agent approaches. Despite this empirical paper, they explicitly note that it remains undetermined which scenarios and dimensions truly favor multi-agent systems, what bottlenecks underlie their performance, and whether multi-agent coordination consistently yields benefits.