Agent Nodes

Introduction to AgentFlow Nodes

In AgentFlow, a workflow is constructed using a series of modular Nodes, each representing a discrete function or operation. These nodes are visually connected on the canvas to define the logic, control flow, and data handling of the entire process.Here’s a brief introduction to the main types of Nodes in AgentFlow:

Basic

• Start Node

  

• Function

Serves as the default and fixed entry point for any workflow-based agent. It is automatically created when a new workflow is initialized.

• Configurable Parameters

Users can define additional custom input variables as needed to support downstream logic.

• Input

Always receives the user's query by default. Also accepts any additional user-defined input variables.

• Output

Passes the received query and input variables to subsequent connected nodes for further processing.

• End Node

  

• Function Serves as one of the possible termination points in a workflow-based agent. It collects and outputs results from preceding nodes for final use, external return, or logging.

• Configurable Parameters Users can define custom output variables to export selected values. These values are typically references to outputs from previous nodes.

• Input Receives data from one or more connected upstream nodes. The data may include intermediate results or values computed earlier in the workflow.

• Output Exports the specified output variables. Only variables explicitly defined by the user will be included in the final result.

• Answer Node

  

• Function Serves as a conversational termination point in a workflow, typically used in Chat Flow scenarios. It outputs a natural language response to the user, allowing dynamic embedding of upstream variables. It cannot coexist with an End Node in the same workflow.

• Configurable Parameters Users can define a response template using natural language, where variables from preceding nodes can be embedded using placeholders .

• Input Receives data from one or more upstream nodes, including all referenced variables intended for insertion into the final response.

• Output Generates a natural language response by resolving variable references within the defined template. The response is returned directly to the user as the final message of the workflow.

Transform

• Code Node

  

  • Function Executes custom Python code as part of the workflow. Used for implementing flexible logic, data processing, or integration tasks that require scripting.

  • Configurable ParametersInput Variables: Passed into the code as function parameters.Output Variables: Returned as a dictionary where keys are variable names available to downstream nodes.

  • Input Receives values from upstream nodes corresponding to the user-defined input variable names. These are passed as arguments to the code function.

  • Output Return a dictionary containing one or more key-value pairs. Each key represents an output variable name, which can be used by downstream nodes.A typical example is as follows：

def main(input_text1: str, input_text2: str) -> dict:
    combined_text = input_text1 + " " + input_text2
    length = len(combined_text)
    return {
    "combined_text": combined_text,
    "text_length": length
}

This function main takes two input strings, input_text1 and input_text2. It concatenates them with a space in between to form combined_text. Then, it calculates the length of this combined string and stores it in text_length. Finally, the function returns a dictionary containing both the concatenated result and its length as output variables.

Control Flow

• Condition Node

  

• • Function

  Directs the workflow based on conditional logic. It evaluates one or more user-defined conditions and determines which downstream branch to execute accordingly.

  • Configurable Parameters

    • A referenced variable (from upstream nodes)

    • A logical expression (e.g., ==, >, in, etc.)

    • An associated output branch (executed if the condition is true) Branches are evaluated in order. Only the first true condition's branch is executed. An optional else branch can be defined as a fallback.

  • Input

  Receives variables from upstream nodes that are used in condition evaluations.

  • Output

  Executes the branch corresponding to the first condition that evaluates to true. If none match and an else branch is defined, the else branch is executed.

• For Each Node

• • Function: Iterates over an input array and repeatedly executes the nodes within a defined loop boundary.

  • Structure:

    • A fixed entry node: For Each Start

    • A loop boundary box: drag any nodes into this area to make them part of the loop body

  • Usage:

    • Drag nodes from the node list into the loop area

    • Set the Input Variable (must be of Array type)

    • Set the Output Variable (must be the output from within the loop scope.)

  • Execution Logic:

    • Automatically iterates through the array

    • Provides two built-in variables to nodes inside the loop:

      • for_each_index: the index of the current item

      • for_each_value: the value of the current item

  • Optional Setting:

    • Parallel Execution: Enable this option for loops without inter-item dependencies to improve performance.

• Variable Assigner Node

• Function: Sets or updates globally accessible variables within the workflow. Supports both assignment and reset operations, enabling dynamic runtime behavior based on changing context.

• Configurable Parameters:

  • Target variable: The name of the global or referenced variable to modify

  • Assignment value: A constant, expression, or variable reference to assign

  • Operation type:

    • assign: Assign a new value to the variable

    • reset: Clear or reset the variable to its default state

  • Multiple variable operations can be configured in one node

• Input: Receives variables or context from upstream nodes as sources for assignment.

• Output: Passes the updated variables to downstream nodes for use in subsequent steps.

• Human Review Node

  

  • Function

  Pauses the automated workflow and sends a message to a designated human reviewer for manual decision-making. Based on the reviewer’s selection, the workflow proceeds along a specified branch.

  • Configurable Parameters

    • Review Message: A customizable message presented to the reviewer to explain the context and request input.

    • Human Review Type: Specifies the review format (e.g., single choice, multi-choice, comment).

    • Options: A list of selectable options, each including:

      • Option Name: A label for the option (e.g., "Approve", "Reject").

      • Option Description: An optional description to help the reviewer understand what the option means.

    • Reviewer Role: Defines which user group or role is authorized to perform the review.

  • Input

  Receives all necessary variables and context from upstream nodes to assist the reviewer in making a decision.

  • Output

  The selected option is returned as an output variable and determines the next step in the workflow, typically branching based on the review result.

AI Models

• LLM Node

  

  • Function

  • Integrates a Large Language Model (LLM) into the workflow to generate natural language responses, perform text understanding, summarization, translation, or other NLP-related tasks based on provided input variables. Enables conversational AI, content generation, or intelligent decision support.

  • Configurable Parameters

    • System Prompt A fixed or semi-fixed instruction that provides context or role setting for the LLM, guiding its response style and behavior (e.g., "You are a helpful assistant.")

    • User Prompt A dynamic prompt template that incorporates input variables to form the actual query sent to the LLM.

    • Model Settings Parameters like temperature, max tokens, top-p, frequency penalty, presence penalty, etc., to control output randomness, length, and style.

    • Input Variables Variables from upstream nodes used to fill placeholders in the User Prompt.

    • Output Variables Variables that store the LLM’s generated responses for downstream processing.

  • Input

  • Receives input variables from previous nodes, which are injected into the User Prompt template to produce the final query for the LLM.

  • Output

  • Returns generated text(s) or structured response(s) from the LLM, mapped to output variables for further use or display.

• Question Classifier Node

• Function: Classifies user input into predefined categories to guide downstream workflow branching based on intent.

• Configurable Parameters:

  • Model: The language model used (e.g., GPT-4o-mini)

  • Chat History: Whether to include prior conversation context

  • Class List: Custom category labels defined by the user; each class maps to an output branch

  • Prompt: The classification prompt used to guide the model's decision. Leaving it blank is also allowed.

  • Input Variables: The input field to be classified, usually the user’s query (e.g., User Query)

• Agent Node

  

  • Function

  • Allows invoking another pre-built agent as a node within the current workflow. This supports modular reuse of existing logic and enables hierarchical or nested workflow designs.

  • Configurable Parameters

    • Agent Reference Select or input the ID/name of the existing agent to be invoked.

    • Input Variables Define the variables to be passed into the referenced agent. These should match the input schema expected by that agent.

  • Input

  • Receives variables from upstream nodes and passes them as input to the specified agent.

  • Output

  • Returns the predefined output variables of the referenced agent. These outputs are automatically exposed for use in subsequent nodes.

Knowledge

Knowledge Search Node

• Function

Retrieves relevant information from a specified knowledge database based on user-defined input variables. Typically used to provide external knowledge support for downstream reasoning or generation.

• Configurable Parameters

  • Knowledge Database Select the knowledge base to search from (e.g., FAQ base, product manual, research corpus).

  • Input Variables Specify the variables whose values will be used as search queries.

  • Top K Set the maximum number of results to return from the knowledge base.

  • Score Threshold Define the minimum similarity score required for a result to be included.

• Input

Receives search query inputs (usually from upstream node outputs or user inputs), which are used to query the selected knowledge base.

• Output

Returns a list of relevant knowledge entries, each typically including fields like content, source, and score. This list is passed to downstream nodes.

Tool

• Tool Node

  

• Function

Enables integration of external tools within a workflow. Users can either select from built-in tools provided by the platform or use custom tools they have created. The Tool Node extends workflow capability with specialized processing.

• Configurable Parameters

  • Tool Selection Choose a tool from the built-in library or link a user-defined tool.

  • Tool-Specific Configuration Configuration options vary depending on the selected tool. Each tool may define its own required input variables, output structure, and runtime parameters.

• Input

Receives input variables required by the selected tool. These are usually mapped from upstream node outputs.

• Output

Returns the output as defined by the tool's specification. Output variables are exposed for use in downstream nodes.