# Enterprise Deep Research: Steerable Multi-Agent Deep Research for Enterprise Analytics

**Authors:** Akshara Prabhakar, Roshan Ram, Zixiang Chen, Silvio Savarese, Frank Wang, Caiming Xiong, Huan Wang, Weiran Yao  
**arXiv:** [2510.17797](https://arxiv.org/abs/2510.17797)  
**Date:** 2025-10-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2510.17797v1/

## Key points

- EDR is a multi-agent system designed to transform unstructured data into actionable insights for enterprises.
- Master Planning Agent for query decomposition
- Specialized search agents for different sources like General, Academic, GitHub, and LinkedIn
- MCP-based tool ecosystem supporting NL2SQL and file analysis
- Visualization Agent for data-driven insights
- Reflection mechanism that updates research direction based on knowledge gaps
- Key feature: Ability to generate final reports by synthesizing running summaries, aggregated sources, code snippets, and steering history.
- Quality assurance checks ensure completeness, structural coherence, query coverage, and adherence to user directives in the final report.
- EDR outperforms state-of-the-art agentic systems on benchmarks such as DeepResearch Bench and DeepConsult without human intervention.
- Tested against popular proprietary deep research assistants and open-source systems using gemini-2.5-pro as the base model with evaluation datasets including DeepResearch Bench, DeepConsult, and ResearchQA.
- Results show EDR excels in report quality metrics like comprehensiveness, instruction-following accuracy, and citation accuracy compared to other systems while consuming fewer tokens than some open-source alternatives.
- Leverages advancements in language models to provide enterprises with an autonomous research assistance tool that can generate structured reports efficiently.
- Adaptive update of research direction based on knowledge gaps makes it a valuable asset for enterprises seeking data-driven insights.

## Summary

is a multi-agent system designed to help enterprises transform unstructured data into actionable insights. The system consists of various components, including a Master Planning Agent for query decomposition, specialized search agents for different sources like General, Academic, GitHub, and LinkedIn, an MCP-based tool ecosystem supporting NL2SQL and file analysis, a Visualization Agent for data-driven insights, and a reflection mechanism that updates research direction based on knowledge gaps. One key feature of EDR is its ability to generate final reports by synthesizing running summaries, aggregated sources, code snippets, and steering history. Quality assurance checks ensure completeness, structural coherence, query coverage, and adherence to user directives in the final report. This interactive document provides transparency, reproducibility, and user-aligned insight generation. show that EDR outperforms state-of-the-art agentic systems on benchmarks such as DeepResearch Bench and DeepConsult without human intervention. The system is tested against popular proprietary deep research assistants and open-source systems using gemini-2.5-pro as the base model. Evaluation datasets include DeepResearch Bench with complex research tasks across various fields, DeepConsult focused on business prompts, and ResearchQA with scientific test questions curated by PhD scholars. Results show that EDR excels in report quality metrics like comprehensiveness,, instruction-following accuracy, as well as citation accuracy compared to other systems. It consumes fewer tokens than some open-source alternatives like langchain-open-deep-research while delivering exceptional performance. Overall, leverages the advancements in language models to provide enterprises with an autonomous research assistance tool that can generate structured reports from diverse sources efficiently. Its ability to adaptively update research direction based on knowledge gaps sets it apart as a valuable asset for enterprises seeking data-driven insights.