AI Agents in Research Automation: A Comprehensive Review
Bhagyashri R. Wankar, Achyut Shinde, Bhavesh Choudhary, Kartik Aswar, Sonali M. Sonavane, Komal R. Jadhav
G H Raisoni College of Engineering & Management, Pune, India
Emails: bhagya.wankar@gmail.com, avshinde018@gmail.com, choudharybhavesh191@gmail.com, kartikaswar03@gmail.com, sonali.sonavane@raisoni.net, komal.jadhav@raisoni.net
Sagar Satish Gaikwad
Vice President, ACTIN Technologies, Pune, India Email: sagar.gaikwad2829@gmail.com
Abstract—Large language models have, somewhat quietly, pushed Retrieval-Augmented Generation (RAG) into the center of many research-oriented applications. The underlying idea is not particularly complicated. Instead of relying only on what a model absorbed during training, a RAG system reaches outward. It retrieves relevant information from external sources and then builds its response around that material. In theory, that sounds sensible enough. However, in practice, many conventional RAG setups still depend on relatively static knowledge bases. These repositories can age quickly, especially in domains where information shifts almost weekly. As a result, the system’s reasoning can feel narrow. Verification steps are often minimal, and incorporating newly emerging knowledge is, at times, more cumbersome than one might expect.
Because of these limitations, researchers have started looking beyond the traditional retrieval pipeline. Recently, attention has moved toward more agentic forms of AI-driven research systems. The ambition here is a little different. Rather than acting as passive tools that fetch documents and summarize them, these systems are designed to behave more like active research assistants. Work on agent-based reasoning, autonomous research workflows, multi- agent coordination, and self-correcting AI mechanisms hints at a more fluid research environment [?]. In principle, such systems attempt to reproduce some aspects of how human researchers actually work. They plan what to investigate, discover information in real time, examine documents more deeply, compare ideas across multiple sources, and occasionally check their own conclusions through iterative verification.
With that backdrop, this review looks closely at the emerging landscape of agentic research systems. The goal is to identify their strengths, the weaknesses that still persist, and the design patterns that appear to be taking shape. Some of these approaches do suggest meaningful improvements for AI-assisted research. Yet the picture is not entirely settled. Questions around reliability, scalability, and real-world deployment remain open, and perhaps they will continue to be for some time.
Index Terms—Agentic AI, Multi-Agent Architecture, Retrieval- Augmented Generation (RAG), Self-Correcting AI, Large Language Models, Information Retrieval, Document Intelligence.
