LLM Agent Node

The LLM Agent node is a sophisticated component within the FlowAI Smart Workflow ecosystem, distinguishing itself from conventional LLM nodes. As an intelligent agent, the LLM Agent possesses the capability to autonomously invoke tools for executing complex tasks, thereby achieving advanced automated processing capabilities.

LLM Agent Node Basic Configuration

Node Properties Explained

1. Node Name: A customizable identifier for the LLM Agent node. Modifying this name will affect how other nodes in the workflow reference it. For instance, changing “LLMAgent1” to “Intelligent Assistant” will alter references from $LLMAgent1.result to $Intelligent Assistant.result. It is advisable to use descriptive names to enhance understanding and facilitate maintenance.

2. Model: Select the desired language model from the dropdown menu. Currently supported options include:

   • GPT-4.1: The latest high-performance model, offering full Agent capabilities (Now fully supported)
   • GPT-4o: Strikes an optimal balance between performance and cost
   • GPT-4o-mini: Offers a balanced approach to performance and cost
   • DeepSeek: The most economical and cost-effective model
   • Other models: Available for selection based on specific requirements

3. Prompt Editing Area: This section is used to compose instructions for the intelligent agent. You can:

   • Utilize variable syntax to reference outputs from other nodes
   • Craft detailed task descriptions
   • Specify expected output formats

4. Output Language: Define the Agent’s response language, with support for multiple options including Chinese, English, and others. (Note that the output quality is contingent upon the LLM’s level of intelligence.)

5. Tools: A suite of tools accessible to the Agent, currently including:

   • Web Scraping: Automates the retrieval of web content
   • Calculator: Performs mathematical computations
   • HTTP Request: Sends API requests to fetch data

6. Maximum Iteration Rounds: Specifies the maximum number of tool calls the Agent can execute, ranging from 1 to 50.

Distinguishing LLM Agent from LLM Nodes

Feature	LLM Agent Node	LLM Node
Functionality	Intelligent agent capable of decision-making and tool use	Single AI inference for text content generation
Tool Calling	Supports autonomous invocation of multiple tools	Does not support tool calling
Task Complexity	Ideal for complex, multi-step tasks	Suited for single text generation tasks
Autonomy	High, with ability to select tools based on task demands	Low, adhering strictly to provided prompts
Application Scenarios	Data collection, analysis, and automated processing	Content generation and text analysis

Prompt Writing Techniques

In contrast to LLM nodes, prompts for LLM Agent nodes emphasize task description and goal definition, rather than detailed execution steps.

Basic Syntax

Please search for the latest information on "$input.keyword" at http://xxxx.com and provide a concise summary.

Multi-step Task Example

Please execute the following tasks:
1. Retrieve the stock price for "$input.stock_code"
2. Calculate the average price over the past 7 days
3. If the average price exceeds $input.threshold, dispatch an HTTP request to our notification system

Note:
The stock query method is as follows:
1. Access https://xxxx.com/stock/xxxx
2. Replace xxx with the actual stock code

Data Processing Example

Please analyze the following dataset: $input.data_table
Requirements:
1. Compute the mean value for each column
2. Identify outliers (data points exceeding two standard deviations)
3. Generate a comprehensive analysis report, including key findings and recommendations

Detailed Tool Usage

The LLM Agent node can leverage multiple tools, each with specific use cases and capabilities:

Web Scraping Tool

The web scraping tool empowers the Agent to automatically extract web content, suitable for:

• Gathering the latest news or information
• Obtaining product prices or specifications
• Extracting structured data from web pages

Example Prompt:

Please scrape the webpage at "$input.url", extract all product names and prices, and organize the information into a Markdown table format.

Calculator Tool

The calculator tool enables the Agent to perform mathematical operations, ideal for:

• Data analysis and statistical computations
• Financial calculations
• Scientific calculations and unit conversions

Currently, the calculation tool supports basic mathematical syntax, including arithmetic operations, trigonometric functions, Pi, exponents, and more.

Example Prompt:

Please calculate the following statistical metrics for this dataset: $input.number_list
Required calculations: mean, median, standard deviation, and quartiles

HTTP Request Tool

The HTTP request tool allows the Agent to interact with external APIs, suitable for:

• Data synchronization and integration
• Triggering operations in external systems
• Retrieving data from third-party services

Example Prompt:

Please transmit the following processing result: $input.processing_result
via an HTTP POST request to our API endpoint: $input.API_address
Ensure that the correct JSON format is utilized.

Advanced Configuration Details

Maximum Iteration Rounds Setting

The maximum iteration rounds parameter controls the number of tool calls the Agent can execute. However, it’s not a strict one-to-one correspondence, as it depends on the AI’s judgment. For certain parallel tasks, multiple tools may be called within a single iteration round.

1. Parameter Specifications:

   • Range: 1-50
   • Default value: Typically 3-5

2. Setting Recommendations:

   • Simple tasks: 3-5 rounds
   • Complex tasks: 10-15 rounds
   • Special scenarios: Adjust as needed

3. Considerations:

   • Excessive rounds may increase execution time and potentially cause the LLM’s context to exceed the AI’s supported context size
   • Insufficient rounds may result in incomplete task execution
   • It is advisable to assess your specific task requirements and select an appropriate number of rounds

Node Output Utilization

The output of the LLM Agent node is accessed via $LLMAgentNodeName.result, for example: $LLMAgent1.result

The output typically encompasses:

• The Agent’s reasoning process
• Records of tool calls
• Final conclusions or processing results

Common Application Scenarios

• Data Collection and Analysis: Automatically gather data from multiple websites, perform calculations and analysis, and generate comprehensive reports
• Automated Customer Service: Autonomously query knowledge bases, calculate prices, or dispatch notifications based on customer inquiries
• Content Aggregation and Processing: Collect information from diverse sources, integrate and analyze to produce summaries or reports
• Intelligent Monitoring System: Regularly inspect data sources, perform calculations and assessments, and trigger appropriate actions
• Automated Workflow: Receive input data, invoke multiple tools for processing, and output structured results

Usage Recommendations and Best Practices

1. Task Definition:

   • Clearly articulate task objectives and expected outputs
   • Decompose complex tasks into manageable steps
   • Provide comprehensive context information

2. Tool Selection:

   • Enable only the tools necessary for the task (as additional tools consume more tokens)
   • Consider the sequence and dependencies of tool calls
   • Provide clear usage guidelines for each tool

3. Performance Optimization:

   • Set an appropriate maximum iteration rounds
   • Avoid overly broad or ambiguous instructions
   • Provide well-structured input data

4. Debugging Techniques:

   • Initiate testing with simple tasks
   • Monitor the Agent’s tool calling process closely
   • Refine prompts based on execution outcomes
   • Document successful cases as templates for future reference

By effectively configuring and utilizing the LLM Agent node, you can achieve highly automated intelligent workflows, significantly enhancing efficiency in handling complex tasks. For more detailed information about variable syntax, please consult the Variable Syntax Guide.

Tip

When implementing the LLM Agent node, it is recommended to begin with simple tasks for testing purposes, familiarize yourself with the Agent’s operational methodology, and then progressively increase complexity. Simultaneously, closely monitor the Agent’s tool calling process to ensure it executes as intended.