TLDR: Enterprise Deep Research (EDR) is a multi-agent AI system developed by Salesforce AI Research to help enterprises transform unstructured data into actionable insights. It features a Master Planning Agent for adaptive query decomposition, four specialized search agents (General, Academic, GitHub, LinkedIn), an extensible tool ecosystem (NL2SQL, file analysis, visualization), and a reflection mechanism that allows for optional human-in-the-loop steering. EDR automates report generation, provides real-time updates, and integrates seamlessly into enterprise workflows. It has demonstrated superior performance on deep research benchmarks and achieved high user satisfaction and efficiency gains in internal enterprise use cases, offering a transparent and adaptable approach to complex business analytics.
In today’s fast-paced business world, companies are constantly bombarded with vast amounts of unstructured information. Turning this raw data into clear, actionable insights is a major challenge. While autonomous AI agents offer significant promise, they often struggle with the specific details of a business domain, understanding user intent, and integrating smoothly into existing enterprise systems.
Addressing these critical limitations, Salesforce AI Research has introduced a groundbreaking system called Enterprise Deep Research (EDR). EDR is a sophisticated multi-agent framework designed to transform how enterprises conduct deep research and generate analytics. It’s built to be transparent, adaptable, and aligned with user needs, allowing for dynamic guidance during the research process.
How EDR Works: A Collaborative AI Ecosystem
EDR operates through a carefully orchestrated system of specialized AI agents, working together to tackle complex research questions. At its core, the system includes:
-
A Master Planning Agent: This central orchestrator breaks down high-level research goals into smaller, manageable tasks. It intelligently adapts its strategy based on the user’s query, identified knowledge gaps, and any steering guidance provided.
-
Four Specialized Search Agents: EDR employs dedicated agents for different information sources: a General Web Search agent for broad content, an Academic Search agent for scholarly publications, a GitHub Search agent for code and technical implementations, and a LinkedIn Search agent for professional profiles and company information.
-
An Extensible Tool Ecosystem: Beyond search, EDR integrates various tools to process and analyze data. This includes an NL2SQL agent to translate natural language into database queries, a File Analysis tool for processing diverse document types (PDFs, spreadsheets, etc.), and a Visualization Agent to generate data-driven insights through charts and graphs. The system is also designed to connect with custom enterprise systems via the Model Context Protocol (MCP).
-
A Reflection Mechanism: After gathering and synthesizing information, EDR has a unique reflection process. It identifies missing knowledge, evaluates task alignment, and detects inconsistencies. This mechanism then updates the research direction, ensuring the system stays on track and addresses all aspects of the query. Crucially, it allows for optional human-in-the-loop steering, enabling users to intervene and guide the research dynamically.
This integrated approach allows EDR to automate report generation, provide real-time streaming updates, and deploy seamlessly within enterprise environments.
Steerable Context Engineering: Human-AI Collaboration
One of EDR’s most innovative features is its “steerable context engineering.” Unlike traditional systems that operate as black boxes, EDR makes its reasoning transparent. Users can see the agent’s internal planning state through a ‘todo.md’ file, which acts as both an execution plan for the agents and a progress tracker for the user. This allows humans to act as “context curators,” dynamically modifying the agent’s direction by adding, canceling, or reprioritizing tasks in natural language.
This real-time steering capability is vital for enterprise settings where domain knowledge evolves rapidly and goals are strategic rather than purely factual. It ensures that the research remains visible, modifiable, and traceable, fostering a dynamic human-AI collaboration that maintains contextual grounding over long research horizons.
Also Read:
- Unlocking Multi-Dimensional Data: A New Approach with LLM Agents and Knowledge Graphs
- Adaptive Search: How Reinforcement Learning Powers Intelligent AI Agents
Performance and Impact
EDR has been rigorously evaluated on open-ended benchmarks such as DeepResearch Bench, DeepConsult, and ResearchQA. It consistently outperforms state-of-the-art agentic systems, even without human steering. For instance, on DeepResearch Bench, EDR achieved an overall score of 49.86, surpassing many proprietary and open-source systems. On DeepConsult, it recorded the highest win rate of 71.57% against baselines.
Beyond benchmarks, internal enterprise use cases have shown EDR achieving over 95% accuracy in SQL generation and execution, 99.9% uptime, and a 98% task completion rate in user studies. Users reported a 4.8/5 satisfaction score and a 50% reduction in time-to-insight for complex analytical tasks.
The researchers have also open-sourced the EDR framework and a dataset of research trajectories (EDR-200), which captures the full reasoning process—search, reflection, and synthesis—to advance future research in multi-agent reasoning applications. You can learn more about the technical details and findings in the original research paper: Enterprise Deep Research: Steerable Multi-Agent Deep Research for Enterprise Analytics.
In conclusion, Enterprise Deep Research represents a significant leap forward in AI-driven enterprise analytics. By combining intelligent tool selection, adaptive planning, and cross-system retrieval with human-guided reflection, EDR offers a powerful, transparent, and scalable solution for businesses seeking to extract actionable insights from their complex data landscapes.
