Salesforce's EDR shows its research plan and lets you edit it mid-run

Enterprise data tends to sprawl across email, reports, databases and code repositories. Answering a hard question usually takes not one fact but many, plus the ability to synthesize hundreds of sources with checkable citations and a line of reasoning someone can follow. Ordinary agents and classic RAG systems are often a poor fit here: they give shallow answers, they are hard to steer once running, and they do not explain which sources their conclusions rest on. The Salesforce AI Research team proposes Enterprise Deep Research (EDR), a multi-agent system that makes deep research steerable and transparent enough for real work inside a company.

The core idea: change course mid-run

EDR's defining feature is a steerable context. The user sees the research plan as a todo.md and can correct course at any moment: add a task, reshuffle priorities, sharpen the focus of the query (say, restrict it to particular sources, or weight recent publications more heavily). The system saves intermediate state, ties every conclusion explicitly to its sources, and holds up over long sessions. It is not a black box.

Who does what inside

• Master Research Agent: the orchestrator. It decomposes the query into subtasks, selects tools, cross-checks what the different agents come back with, resolves contradictions and guards the quality of the final answer.
• Specialized search agents: four fast agents for general, academic, GitHub and LinkedIn search, each tuned to its own domain, with deduplication and sensible source ranking.
• Domain tools over MCP: files and a company's internal systems, connected through purpose-built tools such as document analysis in Confluence or Notion, along with other MCP tools.
• Visualization Agent: it can plot charts and place the visualizations directly into the report.
• Reflection mechanism: emulates reflection, surfaces gaps in what has been learned, and calls for the plan to be refined and for particular aspects of the query to get more attention.

How it works

1. It starts with the user's query, plus any extra materials they choose to upload. The system lays out a visible plan of three to five tasks with priorities and the origin of each task (the original query, a knowledge gap, or a user instruction). Three depth modes are available: fast, standard and deep.
2. The orchestrator generates new subtasks and search queries, decides which specialized agent or domain tool to use, deduplicates, applies constraints (only certain domains, for example), and gathers results under a cap on how many times each tool can be called.
3. The search agents collect answers from the various sources. The system strips duplicates, normalizes links and condenses the several results of a single query. EDR's LLM then takes the short summary of each task from the previous round, the abstract of the uploaded files and the previous round's digest, and produces a new answer. Importantly, it compresses that result while preserving the source links, because over many iterations the context window limits can run out.
4. The user can steer the process gently without pausing the system. Their messages are queued and applied during the next reflection step, which keeps changes atomic and avoids races.
5. At the end of each iteration cycle the system reconciles every task and source in order to spot knowledge gaps, cancel or reopen the tasks that close the questions it still has, add new ones, and track what is already done. Then the whole loop runs again until coverage is sufficient, at which point EDR produces a report with a reproducible trail and links to the sources.

What sets this agent apart from ordinary ones

• It works from a research plan visible to the user and to every agent, tied to the planned tasks and to the sources behind each action.
• It is transparent to the user.
• It is steerable not only before launch but while it runs.
• It has a built-in mechanism for checking consistency and answer quality.
• MCP makes connecting internal databases and other services easy and cheap
• Context is compressed without losing links and other metadata, which matters for long sessions
• Automatic fact-checking and a confidence estimate for the conclusion are built in

What the experiments showed

The authors claim EDR outperforms the best open deep-search agents on the DeepResearch Bench and DeepConsult benchmarks, even with no user intervention. They also tested EDR on Salesforce's internal data. Notably, the researchers released the EDR-200 dataset with complete scenarios for evaluation and fine-tuning, since many deep research agents are closed.

The engineering

The system is built on FastAPI, with streaming, background task queues and detailed errors. There is a React front end using Material-UI and React Query, plus support for a range of LLM providers with authentication. File upload, file analysis and connections to various databases work out of the box.

What this means for us

EDR is useful because it makes agentic systems more reliable and more controllable, which is what a business needs. The user can see the path the system took to any given answer. If they do not like the result, they can redirect the search and get a reproducible report with sources, visualizations and a history of the decisions made along the way. The multi-agent approach and MCP together cover everything from market reviews to code audits to analytics over internal databases. The open code and dataset let you verify the conclusions and keep developing the system without depending on its authors.

So AI does not replace the analyst; it becomes the analyst's partner. A business can now get more than answers — it can understand why those particular answers are right, and make genuinely grounded decisions on that basis.

💾 Code