Understanding Agentic Search Behavior
This study provides a systematic characterization of how conversational LLM agents perform web searches. Unlike traditional keyword-based search, agentic search involves a multi-step process: query formulation, result evaluation, and synthesis. The research highlights that agents do not search uniformly; instead, they employ varying strategies depending on the complexity of the user's intent.
Search Strategies and Decision-Making
The authors identify that an agent's performance is heavily dependent on its 'search strategy'—the iterative process of deciding when to search, how to refine queries, and when to stop gathering information. Key findings include:
- Query Refinement: Agents that perform multiple, iterative searches often yield higher-quality responses, but they are also prone to 'query drift,' where the agent loses focus on the original user intent.
- Information Overload: Agents often struggle to filter noise from search results. The study notes that the quality of the final response is not strictly proportional to the number of search results processed; rather, it depends on the agent's ability to prioritize relevant information over redundant or low-quality sources.
- Decision Thresholds: The research highlights a critical trade-off between search depth and latency. Agents that are configured to be more 'thorough' often produce better answers but significantly increase the time-to-first-token, which can degrade the user experience in conversational interfaces.
From Search Results to Response Synthesis
The study concludes that the final response quality is a function of both the retrieval mechanism and the synthesis capabilities of the LLM. Even with high-quality search results, agents can fail to synthesize information correctly if the prompt instructions are ambiguous or if the model suffers from hallucination during the summarization phase. The authors suggest that future agentic architectures should focus on better 'result evaluation' layers that act as a filter between raw search output and the final generation step to ensure higher factual accuracy.