EDRA: Enterprise Deep Research Agent

Abstract

Many critical enterprise functions—such as strategic sales, market due diligence, competitive analysis, regulatory strategy, and R&D planning—depend on teams conducting research. This often involves manually digging through specialized data sources, piecing together information from different teams and tools, synthesizing fragmented insights into coherent narratives, and doing it all while protecting sensitive or confidential details. To address these challenges, we introduce the Enterprise Deep Research Agent (EDRA), an agentic system designed to support secure, adaptive, and high-fidelity research workflows. EDRA combines a privacy-aware architecture with adaptive reasoning and planning capabilities, enabling enterprises to scale deep research more effectively. We contribute: (1) an autonomous agent capable of leveraging enterprise and public tools across hundreds of action and reasoning steps to generate high-quality research reports; (2) an ablation study analyzing the impact of key architectural components on agent performance; and (3) state-of-the-art performance of 67.2% on the enterprise deep research agent benchmark DRBench. Code available at https://github.com/ServiceNow/deep-research.