Inside Enterprise Deep Research’s steerable investigation

An enterprise research question can span public information, internal files, structured data, and several kinds of analysis. Enterprise Deep Research organizes that work around a master plan that evolves as specialized agents collect evidence.

A plan connects several search surfaces

The Master Planning Agent decomposes a query into research tasks. Four search specialists cover general web information, academic material, GitHub, and LinkedIn. A shared running summary brings their outputs together and helps the system decide which parts of the question need further work.

An extensible Model Context Protocol tool layer adds capabilities such as natural-language-to-SQL queries and file analysis. A Visualization Agent can turn relevant data into graphical explanations. These roles connect information gathering with the analytical work needed for a report.

Reflection updates the investigation

The system records research tasks and revisits them as results arrive. Its reflection mechanism identifies knowledge gaps and updates the direction of the investigation. Users can contribute steering guidance during this process, giving the planner additional context about the intended analysis.

This creates a concrete place for interaction: a person can respond to the developing investigation while the system maintains its tasks, evidence, and accumulated findings.

Reports as a coordinated workflow

The paper evaluates the system on open-ended research benchmarks and describes its report-generation and streaming implementation. The released framework and benchmark trajectories expose how planning, tool use, and synthesis fit together.

The ResearchTodoManager records task identifiers, priorities, and provenance. It distinguishes tasks originating in the initial query, tasks created to address a knowledge gap, and tasks introduced by user steering. A persistent task representation carries those decisions through the investigation, making changes in research direction part of the system’s recorded workflow.

Akshara Prabhakar and colleagues developed this work at Salesforce AI Research. The authors include Weiran Yao and Frank Wang. Read the paper and inspect the implementation.