Skip to main content
SaaS ·
Available on SaaS with FlowX.AI . This feature is live on managed (SaaS) deployments now. Self-hosted deployments will receive it with the next LTS release family.

​
Overview

The Deep Research node investigates a question across the open web and files a cited report into a Knowledge Base. It is a budgeted, multi-pass agent: it plans the work, sends sub-researchers to search and read pages, grades what they bring back, and keeps going until the budget runs out or it has enough to report. Unlike a Custom Agent, you do not write the method. The research prompt is built into the node, so you describe the subject and the node decides how to pursue it.
The result always lands in a Knowledge Base. Research is filed as text so retrieval nodes and other agents can reach it later. A rendered document is optional, and off by default.

​
Adding a Deep Research node

Open your workflow in Integration Designer, then pick Deep Research from the AI Agents group in the left node panel, alongside Custom Agent and Intent Classification Agent.

​
Configuring the node

​
string
required
What the node should investigate. Bind workflow data with ${...} placeholders, and add any standing guidance you want the node to follow.Example:
There is no Instructions field on this node. The research method itself is built in, so this field carries the subject and your guidance, and nothing about how to search.
​
string
How far the node fans out, chosen from four options.Default: StandardQuick, Standard and Deep apply budgets tuned by the platform, shown under the control as a summary such as 24 passes · 6 tool calls per sub-researcher:Custom replaces the summary with two fields you set yourself:
The number of sub-researchers is always the node’s decision. It is not a setting, on any depth.
​
object
Which parts of the open web the node may read. Leave both lists empty to allow anything.Open Manage lists to edit them:
An allow list is a hard restriction. Nothing outside it is read. An over-strict list yields no sources and fails the node.
Each list holds up to 50 entries and takes either mode:
  • Static — type hostnames such as reuters.com. Anything that is not a hostname is rejected, and a duplicate is refused with Already in the list.
  • Dynamic — bind a single ${...} expression that resolves to the whole list, such as ${allowedDomains}. Anything past the first 50 entries is ignored at runtime.
​
object
required
Where the report is indexed so retrieval nodes and other agents can reach it. All three fields are required.
Check the store name. An unrecognised name creates a new store rather than failing, so a typo produces a second store that looks empty instead of an error you can see.
​
boolean
Renders the research as a document in addition to filing it as text, and lets you choose the format and where it lands.Default: falseThe research is always filed as raw text for the Knowledge Base, so turning this off does not lose the result. Turn it on when someone needs a document to read.
​
string
The model the sub-researchers use, set on the AI Model tab.Default: Same as orchestratorA cheaper model here is the usual way to make a wide fan-out affordable, since the sub-researchers do most of the calls.
​
string
required
The key the node’s output lands under in workflow state.Default: responseKey
This node has no Request timeout setting. The shared timeout does not apply to it, so the control is not shown rather than shown and ignored.

Custom Agent node

Agents where you write the instructions and attach the tools

Knowledge Base overview

Where the research is filed, and how to retrieve it
Last modified on September 3, 2026