> ## Documentation Index
> Fetch the complete documentation index at: https://docs.prisme.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Deep Research

> A capability that runs a multi-round, iterative-deepening investigation over your agent's own tools and returns a structured, cited report.

**Deep Research** is a capability you add to any agent. When the agent calls it, it runs a
bounded, multi-round investigation — plan → gather → reflect → iterate → synthesize — **over
that agent's own retrieval tools** (a knowledge base, Bing/Brave/Google search, any search MCP)
and returns a structured, cited report. It's the difference between a single quick lookup and a
thorough, sourced answer.

<Note>
  Deep Research is a **system tool** (like Memory) — drop it onto an agent from the capability
  catalog and it immediately operates over whatever tools that agent already has. It does not
  bring its own data source; it orchestrates the agent's.
</Note>

## What it does

A normal turn does one pass: the model may call a tool once and answer. Deep Research instead
runs a short **sub-loop**, dedicated to one research question, with its own budget:

1. **Plan** — break the question into 3–6 sub-questions.
2. **Gather** — for each sub-question, query the agent's retrieval/search tools (several calls per round).
3. **Reflect** — assess what's answered vs missing or contradictory; turn gaps into new sub-questions.
4. **Iterate** — keep going until coverage is good or the budget runs out; stop early when a round adds nothing new.
5. **Synthesize** — return a structured report: *Executive summary · Findings (with inline citations) · Open questions · Sources*.

The report flags any claim not backed by a cited source, and never fabricates data or sources.

## It adapts to the agent's tools

Deep Research uses **whatever retrieval tools the host agent has** — that's the whole point.

| The agent has…                                                   | What Deep Research does                                                                                                         |
| ---------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------- |
| A web search tool (Bing, Brave, Google) **and** a knowledge base | Plans sub-questions and queries **both** per sub-question, cross-checks, iterates                                               |
| A knowledge base only                                            | Iterative RAG with query reformulation each round                                                                               |
| A web search MCP only                                            | Iterative web search, always citing sources                                                                                     |
| **No retrieval tool at all**                                     | Falls back to multi-pass reasoning + self-critique, and **explicitly flags "no external sources"** — it will not invent sources |

It is **read-only**: it never calls write/create/update/send actions, and it cannot call itself
(recursion-safe).

## Add it to an agent

1. Open the agent in **Agent Creator** → **Capabilities** → **Add Capability**.
2. Add **Deep Research** (category *Research*). It appears as a tool the agent can call.
3. *(Recommended)* also add the **Deep Research** skill — it tells the agent *when* to use deep
   research and how to present the report. Pure guidance; no extra setup.
4. Make sure the agent has **at least one retrieval tool** (a knowledge base or a web search MCP)
   so the research can cite real sources. Without one, Deep Research still runs but answers from
   reasoning only, marked unverified.

<Tip>
  Deep Research uses the **agent's own model** for its sub-loop. If the agent's model is invalid or
  its `max_tokens` is very low, Deep Research will fail like any other turn — check the agent's
  model settings first.
</Tip>

## Budget

The research sub-run has its own budget, independent of the main conversation, configurable on the
capability:

| Setting            | Default | What it caps                                      |
| ------------------ | ------- | ------------------------------------------------- |
| `max_turns`        | 6       | Maximum rounds/turns inside the research sub-loop |
| `token_budget`     | 60000   | Token ceiling for the whole sub-run               |
| `tool_call_budget` | 20      | Maximum tool calls during the research            |

Higher budgets = deeper, more thorough (and slower, costlier) research. These are server-side; the
model cannot raise them itself.

## When to use it — and when not to

**Use Deep Research for**: open-ended or comprehensive questions that need several searches and
synthesis — *"comprehensive analysis of X", "everything about Y", "investigate Z thoroughly",
state-of-the-art reviews, competitive landscapes.*

**Don't use it for**: a single quick fact or a one-shot lookup — a normal turn (or one direct tool
call) is faster and cheaper. The Deep Research skill already instructs the agent to skip it for
quick questions.

## Output

The report is rendered in the conversation (and the canvas for substantial reports), structured as:

```
## Executive summary
## Findings        (per sub-question, with inline [source, date] citations)
## Open questions
## Sources         (deduplicated list)
```

Because it can take several rounds, expect Deep Research to be noticeably slower than a normal
reply — it's doing many searches and a synthesis pass, not a single answer.