Deep Research in the Era of Agentic AI: Requirements and Limitations for Scholarly Research

Abstract

In the fast-evolving era of agentic AI, Large Language Models (LLMs) from major providers and open-source alternatives offer unprecedented capabilities for “deep search”, enabling complex, iterative information retrieval and synthesis crucial for academic endeavors. However, their application in scientific research and paper writing necessitates strict requirements and a critical awareness of inherent limitations, including the risks of unreviewed content, temporal biases, and access barriers such as paywalls. This vision paper discusses a list of requirements that a scientific deep research system should have to become a viable candidate (i.e., to become a valuable system for researchers). As well as a list of limitations that are observed from current systems (industry-grade and community-developed). We also outline a path forward for harnessing agentic AI in scientific discovery and scholarly communication.