Skip to content
Nappai Documentation

Deep Agent

Deep Agent is an agent designed for long assignments with many steps: researching several sources, preparing a report, analysing documents or coordinating several tasks. Before starting, it organises the assignment into a task list, keeps notes and intermediate results in its own working files and can hand parts of the work to subagents, each with its own context.

While it works, it shows a status line in the chat for every tool it uses, so you can see where it is. Use it when a regular Agent falls short because the task is too long or has too many parts.

When to use it

  • Preparing a market report by consulting several sources and tools.
  • Analysing a set of documents and writing a summary with conclusions.
  • Tasks that combine searching, calculating and writing over several steps.
  • Running code in an isolated environment to process data or create charts.
  • Assignments that are best split among specialised agents.

Configuration

ConnectionWhat to connect
Model (required)An AI model component. For long tasks, a model with good reasoning ability works best.
ToolsThe tools the agent can use. You can connect several. With Request Approval the agent can stop to wait for an approval.
Fallback ModelsAppears when you turn on Use Fallback Models in the advanced options. One or more backup models. If the main model fails (the provider does not answer, a usage limit is exceeded, it takes too long or returns an error), the agent repeats that step with the first one and, if that also fails, with the next, in order. The conversation and what the tools already did are kept. A model from a different provider works best.
FieldWhat it does
Agent NameThe agent’s name when another agent uses it as a tool or subagent. Default: DeepAgent.
Agent DescriptionWhat the agent does. Another agent reads it to decide when to turn to it.
User PromptThe template of the message the agent receives. Default: {input}. Each variable in curly braces becomes a field with that name.
System PromptThe agent’s instructions: its role, how it should work and what it should deliver.
InputThe assignment to solve, usually from Chat Input.

Advanced options

FieldWhat it does
Recursion LimitThe maximum number of internal steps in one answer. When it is reached, the agent stops. Default: 1000.
StreamThe answer appears in the chat as it is generated. Turning it on shows Stream Thinking, which also shows the model’s reasoning.
Use memoryThe agent remembers the conversation. Turning it on shows the Memory connection, where you can connect a memory component such as Postgres Memory.
Include State in responseAdds the agent’s internal run data to the answer.
Tool StatusShows a status line in the chat for every tool the agent uses. Default: on. Turn it off, for example, when the Deep Agent works as another agent’s tool and its progress is not of interest.
Enable SandboxLets the agent run code in an isolated environment. Turning it on shows the Sandbox connection, where you connect a component such as Nappai Sandbox.
Enable SubagentsLets the agent delegate tasks to specialised agents. Turning it on shows the Subagents connection, where you connect the Agent output of other agents.
Enable RubricBefore finishing, a reviewer checks the work against your criteria and, if something is missing, the agent keeps working. Turning it on shows the Rubric tab (see below).
Enable SkillsLoads skills into the agent to give it extra knowledge or procedures. Turning it on shows the Skills connection and turns on Use memory. See Skills.
Enable HTML ArtifactsThe agent can publish interactive visual pieces in the chat, such as dashboards, charts or diagrams. Turn it on only for agents with a visual job.

Rubric tab. With Enable Rubric on:

FieldWhat it does
RubricThe criteria that define finished work, one per line. It comes with four example criteria. If you leave it empty, nothing is reviewed.
Grader System PromptThe reviewer’s instructions. It comes pre-filled; if you empty it, the default instructions are used.
Max Grading IterationsThe maximum number of reviews (default 3). If they are reached without approval, the agent delivers its last answer.
Use Separate Grader ModelUses a different model to review. Turning it on shows the Grader Model connection.

Guardrails

In the Advanced tab, Guardrails gives the agent rules it never breaks: every request, and every action before a tool is used, is checked. See Guardrails, with an example video.

Using it as an agent tool

You can connect this component to an agent’s Tools input. In the Tool tab, Tool Name and Tool Description are the name and explanation the agent uses to recognise the tool and decide when to call it: write a clear description of what it does.

Batch processing

With the Batch Mode switch on, the component runs once for each item of the list you connect to Batch Input and returns all the results together as a list. Parallel Executions sets how many items are processed at the same time (1 to 5). Skip if Empty Batch Input skips the component when the list arrives empty, and Skip Branch if Empty Output stops the following components when the result is empty. More in Key concepts.

Output

  • Response: the final result of the work as a message. Connect it to Chat Output or to a component that stores or sends the text.
  • Agent: the configured agent, so another agent can delegate to it.
  • Tool: the agent turned into a tool, for another agent’s Tools input.

Example

The consultancy Horizonte Retail prepares a trends report for its clients every month. Its Deep Agent has a web search tool connected to Tools and these instructions in System Prompt: “Research last month’s retail trends, compare at least three sources and deliver a report with a summary, the main trends and recommendations.” The agent organises the work into tasks, consults the sources, keeps its notes and delivers the report in Response. With Enable Rubric on, a reviewer checks that the report meets the criteria before it is delivered.

Good practices

  • Describe the result you expect in System Prompt (structure, length, language); the agent plans from there.
  • Long tasks use more model calls: turn on only the options the assignment needs.