Title: Enterprise Deep Research (EDR): How Steerable Multi-Agent Systems Are Redefining AI-Powered Research
Meta Description: Discover how Salesforce’s Enterprise Deep Research (EDR) framework uses steerable multi-agent AI to transform enterprise research, enabling real-time human guidance and superior benchmark performance.
Introduction: When Research Agents Learn to Take Directions
In October 2025, Salesforce AI Research open-sourced Enterprise Deep Research (EDR)—a multi-agent system that accepts real-time human guidance during research execution. This isn’t just another “AI research assistant” but an intelligent partner that understands natural language commands like “focus on peer-reviewed sources” or “ignore outdated information.”
Imagine having a tireless research team that automatically decomposes complex questions, searches across diverse sources, and immediately adjusts strategy when it veers off course. This is the revolutionary capability EDR delivers. On the DeepResearch Bench, EDR achieved a comprehensive score of 49.86, outperforming most proprietary systems while using only one-quarter the computational cost of comparable open-source solutions.
Core Innovation: From “Black Box” to Transparent, Steerable Research
The Steering Mechanism: Preventing AI Research From Going Down the Wrong Path
What’s the biggest pain point with traditional research agents? Once launched, they’re like rockets fired into space—you can’t adjust their trajectory mid-flight. If the agent misunderstands your intent or gets stuck in unreliable sources, your only option is to start over.
EDR’s Steering Framework solves this fundamental problem. Through a shared todo.md interface that exposes the system’s internal planning state, users can adjust research direction using natural language in real-time:
# Example: User Intervention Process
User Query: "Analyze ethical implications of AI in healthcare"
-> System Generates Initial Plan: [1. Technical ethics, 2. Patient privacy, 3. Regulatory policies]
User Steering Command: "Prioritize European regulations, ignore US cases"
-> System Immediately Adjusts: [1. European regulations (priority 10), 2. Technical ethics (priority 7), Cancel US-related tasks]
Helpful Analogy: If traditional research agents are like self-driving cars (where you can only watch passively after setting the destination), EDR is more like a race car with active steering—allowing a professional driver (the user) to take control at critical turns to ensure optimal routing.
Multi-Agent Collaboration: A Specialized Research Department
EDR isn’t a single agent but a coordinated “research department” composed of five specialized agents:
-
Master Planning Agent: Principal researcher responsible for problem decomposition and coordination -
Four Search Specialists: General web, academic, GitHub, and LinkedIn search -
Visualization Agent: Expert at transforming data into charts and graphs -
Reflection Mechanism: Quality control and gap identification specialist
This division of labor resembles project teams in large consulting firms, where each member focuses on their area of expertise while coordinating for maximum impact.
Technical Architecture: Enterprise-Grade Research Engineering
EDR’s workflow can be understood through this visual representation:
graph TD
A[User Query] --> B[Master Planning Agent Decomposes Task]
B --> C[Generate Initial Todo Plan]
C --> D{Loop Execution}
D --> E[Task-to-Query Transformation]
E --> F[Parallel Search Execution]
F --> G[Result Aggregation & Deduplication]
G --> H[Incremental Synthesis]
H --> I[Reflection & Todo Update]
I --> J[User Steering Intervention?]
J -- Yes --> K[Apply Steering Commands]
J -- No --> L[Continue Next Cycle]
K --> M[Update Todo Plan]
M --> D
L --> D
D -- Complete --> N[Final Report Generation]
style A fill:#e1f5fe
style N fill:#c8e6c9
style J fill:#fff3e0
Chart Description: EDR employs an iterative research process where each cycle includes planning, execution, synthesis, and reflection phases, with user steering interventions possible at any point to adjust research direction.
Performance: Let the Data Speak
EDR’s effectiveness is confirmed by its performance on authoritative benchmarks:
| Benchmark | EDR Score | Top Competitor | Key Advantage |
|---|---|---|---|
| DeepResearch Bench | 49.86 | 50.62 (WebWeaver) | Instruction Following (50.03), Cost Efficiency |
| DeepConsult | 71.57% Win Rate | 66.86% (WebWeaver) | Business Scenario Adaptation |
| ResearchQA | 68.52% Coverage | 75.29% (Sonar) | General Question Handling |
Notably, EDR performs even better in real enterprise scenarios: 95% SQL generation accuracy, 99.9% system availability, 4.8/5 user satisfaction, and 50% reduction in time-to-insight for complex analytical tasks.
Comparative Analysis: EDR vs. Traditional Research Agents
Compared to systems like OpenAI DeepResearch and Gemini Deep Research, EDR’s differentiated advantages include:
-
Transparency: Traditional systems operate as “black boxes” while EDR provides complete research trajectory tracing -
Adaptability: Static planning versus dynamic, todo-driven adjustment approach -
Enterprise Integration: MCP tool ecosystem supports enterprise-specific workflows like NL2SQL and file analysis -
Cost Control: 53.9M tokens versus 207M tokens (langchain-open-deep-research)
Analogy: If traditional agents are like security cameras filming fixed scenes, EDR is a movie camera operated by a professional cinematographer—able to adjust composition, focus, and angles based on the director’s real-time instructions.
Future Outlook: The Next Evolution of Enterprise AI Research
Speculative Perspective: The following projections based on current technology trends represent possibilities rather than certain predictions.
Short-term (2026-2027): Steering mechanisms will become standard features in enterprise AI agents. EDR’s open-source release will spawn specialized variants optimized for vertical sectors like finance, healthcare, and legal.
Medium-term (2028-2029): We may see “predictive steering”—where systems anticipate user needs and proactively suggest research direction adjustments. Agent collaboration will become more sophisticated, forming cross-organizational “research networks.”
Long-term (2030+): 【Bold Projection】 Deep research agents may evolve into corporate “strategic decision cores,” not only answering “what happened” but predicting “what might happen” and recommending “what to do.” These systems will deeply integrate enterprise knowledge graphs, real-time data streams, and external intelligence, serving as CEOs’ “digital chief of staff.”
Conclusion: Redefining Human-AI Research Collaboration
EDR’s significance isn’t about creating another AI that can write research reports, but about establishing a new paradigm for research collaboration—where humans are no longer just task initiators and result recipients, but active participants in the process and co-owners of quality.
In an age of information overload, value comes not from acquiring more information but from how efficiently we filter, synthesize, and apply it. Through its steerable multi-agent architecture, EDR enables enterprises to combine human strategic thinking with AI’s processing scale—the fundamental reason it stands out in the competitive AI research landscape.
With the release of the EDR-200 dataset and the open-sourcing of the system, we stand at an inflection point: enterprise deep research is no longer exclusive to large corporations but is becoming standard capability accessible to every organization. This represents not just technological progress but a significant step toward democratizing knowledge work.
