Skip to main content

Deep Research

Automate with agents runs a pipeline you already know the shape of. Deep Research is the open-ended case: you hand an agent a question instead of a plan, and it decides what to query, join, and cross-reference across your connected data to answer it — then reports back with its reasoning and the sources it used.

Use it when you don't yet know which tables or steps the answer needs — "which of our funded projects lost their lead maintainer this year?", "where does our dependency graph concentrate risk?" — and you want a grounded, auditable answer rather than a single query result.

Run it in the app​

The in-app Agents page is the quickest way in. Open Agents in your workspace sidebar, click Dispatch, and describe the question in natural language; the agent works in the background against the data your org has connected. See the Quickstart for the dispatch flow and screenshots.

note

The in-app Agents feature is available to Pro and Enterprise tiers. Self-serve users can run the same kind of research through their own agent over MCP — see below.

Run it from your own agent​

Connect your agent over MCP (setup) and give it the research question directly. It works the data with the same tools the rest of this path uses — execute_sql to query, dataset discovery to find what's available — but it chooses the queries. This is the Automate capstone pointed at an open question instead of a fixed pipeline, so the same verify-in-loop discipline applies: the agent checks each result before building on it.

Frame the question so the answer is trustworthy​

  • Scope it to connected data. The agent can only reason over what your org has ingested or can query. Connect the sources first (see Ingesting data).
  • Ask for the sources. Have the agent cite the tables and queries behind each claim, so you can trace a number back to where it came from. A research result you can't audit isn't finished.
  • Prefer specific over sweeping. "Rank our collections by contributor growth last quarter" gets a cleaner, checkable answer than "tell me about our data."

Turn it into something shareable​

A research run is most useful when it lands as an artifact others can open. Have the agent capture its queries and findings in a notebook, then publish and share it — the citations and queries travel with it, so a reader sees both the answer and how it was reached.