Agentic Search in the Wild: Intents and Trajectory Dynamics from 14M+ Real Search Requests

Authors: Jingjie Ning, João Coelho, Yibo Kong, Yunfan Long, Bruno Martins, João Magalhães, et al. (8 authors) Published: 2026-01-24 | Source: arXiv:2601.17617 | Dataset: Hugging Face logs

Overview

LLM-powered search agents are increasingly used for multi-step information seeking, yet the IR community lacks empirical understanding of how agentic search sessions unfold and how retrieved evidence is reflected in later queries. This paper provides the first large-scale log analysis of real agentic search traffic, based on 14.44M search requests (3.97M sessions) collected from DeepResearchGym, an open-source search API accessed by external agentic clients.

Methodology

• Sessionization of raw request logs into search sessions.
• LLM-based annotation of session-level intents and step-wise query-reformulation labels.
• Context-driven Term Adoption Rate (CTAR): a proposed metric quantifying whether newly introduced query terms are lexically traceable to previously retrieved evidence.

Key Findings

• Session structure: over 90% of multi-turn sessions contain at most ten steps; 89% of inter-step intervals fall under one minute.
• Intent-dependent behavior: fact-seeking sessions exhibit high repetition that increases over time, while sessions requiring reasoning sustain broader exploration.
• Evidence-driven reformulation: on average, 54% of newly introduced query terms appear in the accumulated evidence context, with additional traceability to earlier steps beyond the most recent retrieval.

Implications

The behavioral signals identified support:

• Repetition-aware stopping rules for search agents;
• Intent-adaptive retrieval budgeting (allocating more exploration budget to reasoning tasks);
• Explicit cross-step context tracking in agent memory design.

The anonymized logs are publicly released, enabling reproducible follow-up research on agentic search dynamics.

Notes for Practitioners

• DeepResearchGym traffic reflects external *agentic clients* rather than human searchers, so patterns (short sessions, sub-minute intervals, high repetition) characterize machine-driven search loops.
• CTAR offers a lightweight, lexical proxy for measuring whether an agent grounds its next query in retrieved evidence — useful for debugging RAG-style pipelines.
• Candidate engineering uses: early-stopping heuristics, budget control, and detecting degenerate repetitive querying.

References

• Paper: https://arxiv.org/abs/2601.17617
• Dataset: https://huggingface.co/datasets/cx-cmu/deepresearchgym-agentic-search-logs