Deep Research Agent · Use case
How can we run a deep research agent on-premises, on our own infrastructure?
A deep research agent takes a research question, proposes a plan, searches the web in rounds and writes a report in which every claim carries its source. The model, the plan and the report stay on your own infrastructure. Only the search queries go out, to the search service you choose.
Gathering before writing.
And questions that give you away.
- ✗ 01Analysts gather sources by hand before they can write. Public research assistants do the gathering, but the question itself often reveals what the organisation is working on.
- ✗ 02A report whose claims cannot be traced to a source cannot be used in a decision.
Where it stops for a person
After the plan, before the first search.After the plan and before the first search. Nothing is searched until a person approves or edits the plan.
It plans first.
You approve the plan.
Step · 01
A user states the question and any limits on sources.
Step · 02
The agent proposes a research plan and waits for the user to approve or edit it.
Step · 03
It searches in rounds, reads what it finds and decides what to look for next.
Step · 04
It writes a report and cites a source for each claim.
Step · 05
The plan, the searches and the sources are kept with the report.
Reads from · connectors
The open web. Internal documents can be added by connecting a source such as SharePoint or Confluence.
Acts through · MCP servers
A web search server. The agent connects to Tavily, Exa, Brave Search or Firecrawl through MCP servers to search and to read pages.
What you set up
- Choose a search service and connect it.
- Set which sites or kinds of source the agent may use.
- Write test questions to score reports before release.
Where it runs
The agent, its knowledge store and the model run on infrastructure you own or control. You choose the model and serve it yourself, so questions, documents and answers are not sent to a public AI service. The exception is web search: the queries go to the search service you choose.
How you get it today
This page is labelled Use case: the agent is built in Katonic Studio. You describe the job, connect what is named above, add test questions and release the agent once it passes them. When the Catalog holds a tested template for this job, this page will carry the Template label and show how to take it.
Does anything leave our network?
Only the search queries, which go to the search service you choose. The model, the plan, internal documents and the report stay on your infrastructure.
Can it run with no internet connection?
Not for web research. Disconnected, the same agent can research only the internal documents you connect.
How do we check a report?
Every claim cites its source, and the plan and the searches are kept with the report.
Which search services does it use?
The agent connects to Tavily, Exa, Brave Search or Firecrawl. You choose which one.