PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Akshara Prabhakar, Roshan Ram, Zixiang Chen, Silvio Savarese, Frank Wang, Caiming Xiong, Huan Wang, Weiran Yao

arXiv:2510.17797v2cs.CLcs.AI

TL;DR

Enterprise research can require synthesizing hundreds of heterogeneous documents, while deep research systems may lack transparency and real-time steering. EDR combines multi-agent research with context-level user steering and iterative synthesis. The paper reports state-of-the-art performance on deep research benchmarks and internal enterprise evaluations.

  • Problem

    Enterprise research requires synthesizing heterogeneous information, while existing deep research systems can be opaque and inflexible, with limited real-time steering.

    Source: §1 · p. 2

  • Method

    EDR is a transparent, steerable multi-agent framework that lets users modify agent context during execution and iteratively synthesizes new research into a running summary.

    Source: §1 · p. 2, §3.3 · p. 7

  • Results

    EDR achieves state-of-the-art performance on deep research benchmarks and internal enterprise evaluations.

    Source: §6 · p. 13

  • Takeaways & Limitations

    EDR demonstrates the potential of AI-powered automation for enterprise research and decision support by integrating advanced reasoning with enterprise-grade system design.

    Source: §6 · p. 13

    The authors identify improving output factuality through better citation and evidence grounding as future work.

    Source: §6 · p. 13

Abstract

from arXiv · show

As information grows exponentially, enterprises face increasing pressure to transform unstructured data into coherent, actionable insights. While autonomous agents show promise, they often struggle with domain-specific nuances, intent alignment, and enterprise integration. We present Enterprise Deep Research (EDR), a multi-agent system that integrates (1) a Master Planning Agent for adaptive query decomposition, (2) four specialized search agents (General, Academic, GitHub, LinkedIn), (3) an extensible MCP-based tool ecosystem supporting NL2SQL, file analysis, and enterprise workflows, (4) a Visualization Agent for data-driven insights, and (5) a reflection mechanism that detects knowledge gaps and updates research direction with optional human-in-the-loop steering guidance. These components enable automated report generation, real-time streaming, and seamless enterprise deployment, as validated on internal datasets. On open-ended benchmarks including DeepResearch Bench and DeepConsult, EDR outperforms state-of-the-art agentic systems without any human steering. We release the EDR framework and benchmark trajectories to advance research on multi-agent reasoning applications. Code at https://github.com/SalesforceAIResearch/enterprise-deep-research and Dataset at https://huggingface.co/datasets/Salesforce/EDR-200

1 Introduction

Enterprise deep research requires synthesizing heterogeneous evidence while goals and information evolve, but existing systems can be opaque and difficult to steer during execution. EDR presents a transparent, steerable multi-agent framework and contributes an extensible architecture, todo-driven human guidance, and the EDR-200 trajectory dataset.

1 Introduction

  • A complete enterprise research answer may require synthesizing hundreds of heterogeneous documents, rather than locating a single fact.Enterprise information is dispersed across sources, and tasks can demand contextual reasoning, long-horizon planning, and transparent evidence attribution.Source: §1 · p. 2
  • Existing deep research systems are described as opaque and inflexible, lacking real-time steering and often requiring costly restarts when they diverge from user intent.The paper associates this rigidity with redundant API calls, inefficient exploration, and results that deviate from user intent [53] [29] [30].Source: §1 · p. 2
  • EDR is a transparent, steerable multi-agent framework that lets users modify agent context during execution through task additions, cancellations, and reprioritizations.Its design exposes planning state through todo.md and tracks tasks and provenance.Source: §1 · p. 2
  • The paper contributes a modular enterprise architecture, todo-driven human guidance during research, and EDR-200, a dataset of complete research trajectories.The trajectories expose the search, reflection, and synthesis process for analysis of planning and decision dynamics.Source: §1 · p. 2

2 System Architecture

EDR uses a modular multi-agent architecture that combines central planning, structured task management, specialized retrieval, and enterprise connectors. Its Master Research Agent decomposes queries adaptively, replans with new information, and coordinates agent results.

2 System Architecture

  • EDR combines a central planning agent, structured task management, specialized retrieval tools, and extensible enterprise connectors for iterative research.The architecture is modular and designed for scalable orchestration of complex research workflows.Source: §2 · p. 3, Figure 1 · p. 3

2.1 Master Research Agent

  • The Master Research Agent decomposes objectives into tasks, classifies intent, extracts entities, and annotates tasks with priorities, tools, metadata, and dependencies.Decomposition uses the user query, prior knowledge gaps, and steering directives; the agent can dynamically replan as results or directives change.Source: §2.1 · p. 3
  • For complex queries, the agent creates parallel tasks with independent search strategies and uses duplicate-awareness to reduce redundant work.The agent also validates consistency, compares agent results, fact-checks, scores confidence, and aggregates outputs while preserving provenance.Source: §2.1 · p. 3

2 System Architecture

EDR exposes task state for human oversight and connects specialized search and domain tools through an integrated workflow. Its tool ecosystem spans web, academic, code, file, database, visualization, and enterprise-system tasks.

2 System Architecture

  • Temporal markers ground recency-sensitive queries, while uploaded documents are treated as authoritative and prompt complementary searches for external validation or extension..Source: §2 · p. 4

2.2 Research Todo Manager

  • The Research Todo Manager makes task IDs, descriptions, priorities, lifecycle states, and provenance visible in a persistent shared todo.md representation.Users can issue natural-language directives that become task adjustments, exclusions, or priority changes; the todos support both agent planning and human oversight.Source: §2.2 · p. 4

2.3 Search API and Tools

  • Specialized search tools cover broad web content, scholarly publications, code repositories, and professional profiles with domain-specific retrieval and metadata.Academic search supports temporal weighting; GitHub search uses repository-level deduplication, while LinkedIn search is restricted to linkedin.com.Source: §2.3 · p. 4

2.4 Domain-Specific and MCP Tools

  • Domain tools analyze uploaded files, translate natural-language questions into validated SQL, generate data visualizations, and connect custom enterprise systems through MCP.MCP supports HTTP and stdio transport and can connect remote computation services and organization-specific tools.Source: §2.4 · p. 5

2.5 Integration and Workflow Coordination

  • EDR tightly couples task decomposition, search execution, and domain-tool use in iterative loops, consolidating outputs into a running summary and reflection process.These outputs inform the Research Todo Manager and Master Agent, supporting context coherence and cumulative knowledge integration.Source: §2.5 · p. 5

3 Research Flow Mechanism

EDR turns a user query into a structured report through repeated planning, agent execution, synthesis, steering, and reflection. Users can track progress and guide the work as knowledge gaps and priorities change.

3.1 User Query Ingestion and Todo Initialization

  • The ResearchTodoManager starts with a lightweight 3–5 task plan and tracks task priority and provenance, distinguishing initial-query, knowledge-gap, and steering-derived work.Steering-derived tasks receive the highest priority, followed by original-query tasks and knowledge-gap tasks.Source: §3.1 · p. 6

3.2 Task-Query Transformation

  • At each iteration, the LLM uses the original objective, refined summary of prior findings, unresolved information needs, and active steering constraints to adaptively generate 3–7 new tasks and corresponding search queries.Source: §3.2 · p. 6

3.3 Result Aggregation and Incremental Synthesis

  • Agents’ results are deduplicated, synthesized with the prior summary, new research, reflection gaps, and uploaded knowledge, then tracked for citation management.Unused sources are logged for transparency but excluded from final citations.Source: §3.3 · p. 7
  • The synthesis process supports sessions with 10+ iterations and hundreds of sources while refining knowledge across loops.It compresses findings while preserving citation links and metadata instead of accumulating raw search outputs.Source: §3.3 · p. 7

3.4 Steering Integration

  • Steering messages are queued during execution, processed atomically between iterations, and translated into priorities, exclusions, or focus directives.Messages arriving during reflection are preserved for the post-reflection queue.Source: §3.4 · pp. 7–8

3.5 Reflection and Todo Update

  • After each iteration, reflection checks for knowledge gaps, task misalignment, and quality inconsistencies, then adds, cancels, or completes tasks.Unaddressed steering messages remain queued for subsequent loops.Source: §3.5 · p. 8

3.6 Iterative Refinement and Loop Continuation

  • The planning, execution, steering, aggregation, and reflection cycle repeats until gaps are resolved, loop limits are reached, or the report is deemed sufficiently complete..Source: §3.6 · p. 8

3.7 Final Report Generation and Validation

  • The final report combines the running summary, sources, code snippets, and steering history, with checks for citation completeness, coherence, query coverage, and directive adherence.The delivered document preserves an account of the research process.Source: §3.7 · pp. 8–9

4 Evaluations

EDR is evaluated on open-ended research benchmarks and internal enterprise use cases, with steering disabled for fixed-query benchmark runs. Results cover benchmark comparisons, enterprise analytics, and a released collection of full research trajectories.

4.1 Experimental Setup

  • Benchmark runs use Gemini-2.5-pro, cap research loops at five for DeepResearch Bench, ten for DeepConsult, and two for ResearchQA, and disable real-time steering for fixed queries.The settings involve no human intervention.Source: §4.1 · p. 9

4.2 Results

  • EDR achieves an overall score of 49.86 on DeepResearch Bench, outperforming all proprietary and most open-source agentic systems, while consuming 4x fewer tokens than langchain-open-deep-research.Source: §4.2 · p. 10
  • 71.57% win rate, 6.82 average score, and 9% lose rate: EDR leads other open-source systems in pairwise comparison against OpenAI DeepSearch.One EDR run uses 48,850,862 tokens and costs $105.51.Source: §4.2 · p. 10, Table 2 · p. 11
  • On ResearchQA, EDR performs strongly on General, Impact, and Comparison items, but has citation failures, example-generation weaknesses, and lower performance on some multi-criteria and domain combinations.Citation failures are reported at 85%; rubric scores are strongly bimodal, and Humanities & Arts struggles with examples.Source: §4.2 · p. 11
  • 75.3% overall coverage for Perplexity Deep Research (Sonar) versus EDR’s 68.5% on ResearchQA, where EDR ran for at most two research loops.Table 3 reports coverage across seven research domains.Source: Table 3 · p. 11, §4.2 · p. 11
  • Internal enterprise evaluations report over 95% SQL-generation and execution accuracy, 99.9% uptime, 98% task completion, 4.8/5 satisfaction, and 50% less time-to-insight.Evaluations include open-ended research and research over complex internal proprietary databases.Source: §4.2 · p. 11

5 Related Works

The related-work discussion identifies a gap in transparent, steerable deep research over heterogeneous proprietary enterprise data. EDR addresses it with adaptive planning, specialized retrieval, and reflection, while its trajectory collection supports analysis of long-horizon research behavior.

5 Related Works

  • EDR-200 trajectories expose search, reflection, and synthesis; report synthesis peaks at iterations 4–5 with a +1,785-word gain, while reflection often identifies market, comparative, and cost-analysis gaps.Across 1,422 reflection instances, the recurring gap shares are 27.1%, 18.4%, and 14.2%, respectively; source usage remains about 14 per iteration.Source: §5 · p. 12, Table 4 · p. 12
  • Rigid draft-then-retrieve pipelines fix outlines before gathering evidence, a design linked to incoherence and hallucination [11].Source: §5 · p. 12
  • Existing systems cannot yet combine enterprise-scale deep research over proprietary, heterogeneous sources with interpretability and steerability.The sources include databases, internal reports, and document repositories.Source: §5 · p. 12
  • EDR combines adaptive planning, specialized retrieval, and human-guided reflection in a transparent, auditable, domain-aware research framework.Its reasoning is externalized into a modifiable plan for human control and auditability.Source: §5 · pp. 12–13

6 Conclusion

EDR is presented as a multi-agent research framework for enterprise analytics, combining adaptive planning, tool selection, cross-system retrieval, and steerable context engineering. The conclusion reports strong benchmark and internal evaluation results while identifying factuality and integration as future-work areas.

6 Conclusion

  • EDR combines intelligent tool selection, adaptive planning, and cross-system retrieval for large-scale, transparent, goal-aligned research workflows.The framework is introduced as a multi-agent autonomous research system advancing enterprise analytics through steerable context engineering.Source: §6 · p. 13
  • EDR achieves state-of-the-art performance on deep research benchmarks and internal enterprise evaluations.The conclusion describes these results as evidence of effectiveness in complex analytical environments.Source: §6 · p. 13
  • Future work targets output factuality through improved citation and evidence grounding, predictive steering mechanisms, and broader enterprise-data integration.Source: §6 · p. 13

A Prompts

The appendix presents prompts in invocation order, covering initial task planning, query decomposition and execution, and post-synthesis reflection. Together, they guide an iterative research loop that identifies gaps and updates the task queue.

A Prompts

  • The initial task-decomposition prompt generates a plan of 3–5 prioritized tasks to guide the first research loop.The appendix lists this prompt first in the sequence.Source: Figure 4 · p. 17, A Prompts · p. 17
  • The task-to-query prompt generates precise, context-aware search queries from research tasks, aligned with user-provided knowledge.Source: A Prompts · p. 17
  • The reflection prompt audits coverage, identifies missing knowledge, updates the task queue, and iterates until critical gaps are closed.It can mark tasks completed or cancelled, or add new search tasks.Source: A Prompts · p. 17

B.1 API Overview

The REST API exposes research, file, database, and interactive-steering capabilities. It also supports real-time streaming and asynchronous task processing.

B.1 API Overview

  • The API exposes comprehensive research through POST /deep-research, with real-time updates available through a streaming endpoint.Source: §B.1 · pp. 17–18
  • File endpoints support individual and batch uploads, analysis, and retrieval of original file content and analysis results.The API also includes status checks, listing, and deletion endpoints for uploaded files.Source: §B.1 · p. 19
  • Database endpoints support SQLite, CSV, and JSON uploads, database listing and schema access, natural-language queries, and deletion.Source: §B.1 · p. 19
  • Steering endpoints accept natural-language messages and expose plans, status, interactive sessions, session listings, and examples.Source: §B.1 · p. 19
  • Advanced API features include SSE updates, asynchronous background processing, error handling, generated API documentation, and configurable rate limiting.The passage also lists CORS support, cache control, and multiple LLM-provider authentication methods.Source: §B.1 · p. 19

B.2 Frontend Architecture

The frontend uses React and TypeScript alongside Material-UI and Tailwind CSS. React Query handles data fetching and state synchronization, including caching and retries.

B.2 Frontend Architecture

  • The frontend uses React 18 and TypeScript, with concurrent rendering, automatic batching, and strict type checking.Source: §B.2 · p. 19
  • Material-UI provides accessible, responsive components, while Tailwind CSS supports responsive styling, dark mode, and accessibility utilities.The interface also includes sortable tables, validated forms, and interactive visualizations.Source: §B.2 · pp. 19–20
  • React Query manages data fetching and state synchronization through caching, background refetching, optimistic updates, and automatic retries.Source: §B.2 · p. 20

Related papers

Loading 2510.17797v2…