Deep Agent

The Deep Agent is an advanced-level AI orchestration component, designed to solve complex, multi-step tasks that require planning, long-term information management, and the delegation of tasks to other specialized sub-agents.

This agent goes beyond a simple question-answer cycle, incorporating a virtual file system to manage information and the ability to act as a supervisor over a team of sub-agents.

Main Use Cases

Use the Deep Agent when you need to build complex systems such as:

Autonomous Research Agents: Capable of planning a research project, searching for information, saving it in temporary files, analyzing it, and drafting a final report.
Orchestration Systems (Supervisor): Where a main agent receives a task and delegates it to the most suitable sub-agent from a team of specialists.
Automations with State Management: For flows that need to create, read, and modify files or data in a virtual “hard drive” throughout their execution.

Configuration (Common Tab)

These parameters define the agent’s identity and fundamental behavior.

Parameter	Technical Description	Purpose in the Flow
`Agent Name`	A descriptive name for the agent.	Used to identify the agent, especially when exposed as a tool for another agent.
`Agent Description`	A clear description of the agent’s capabilities and purpose.	It is crucial when used as a tool, as a supervisor agent will read it to decide when to delegate a task.
`User Prompt`	The template that formats the user input before being processed.	Allows for adding context or structuring the question. The `{input}` variable is replaced with the user’s text.
`System Prompt`	The main instruction defining the role, rules, and work methodology of the agent.	Acts as the agent’s “constitution.” For a Deep Agent, this prompt usually includes instructions on how to plan and use its file system.
`Input`	The direct data input that triggers the agent’s execution.	This is the main field where the user’s request or the objective to be solved is connected.
`Tools`	Connection for components the agent can use as external capabilities.	Connect tools like `Web Search`, `API Request`, etc., here.
`Model`	Connection to the Language Model (LLM) that powers the agent’s reasoning.	The main “cognitive engine.” Connect a component like `Gemini` or `OpenAI` here.
`Sub-agents`	Connection for other compiled agents this agent can supervise and delegate tasks to.	Allows creating hierarchical architectures where this agent acts as a “Director” orchestrating a team of specialist agents.
`Skills`	Connection for predefined functionalities loaded into the agent’s file system.	Enriches the agent with ready-to-use complex capabilities (e.g., a skill to analyze code, another to write reports).

Configuration (Advanced Tab)

These parameters allow for detailed control over performance, output, and the agent’s unique capabilities.

Execution Control

Max Iterations: Sets the maximum number of reasoning cycles the agent can perform. This is a crucial security measure to prevent loops and control resource consumption in complex tasks.

Output

Stream: Allows the agent’s response to be sent word by word, improving the user experience in chat interfaces.
Include State in Response: If activated, the agent’s output will include the full internal state, which is very useful for advanced debugging.
Use Structured Output: Forces the agent to generate its final response in a strict JSON format, defined by a schema.

Memory and Internal Capabilities

Use Memory: Activates conversational memory so the agent remembers past interactions within the same session.
Enable Skills: Allows the agent to use the connected Skills. If disabled, the agent will ignore the skills even if they are connected.
File System Token Limit: Defines the maximum size of the agent’s virtual “hard drive.” This is the token threshold before the system begins discarding the content of the oldest files to manage context.