Agent Node

The Agent node leverages AI language models to generate text, analyze content, answer questions, and transform data using natural language processing.

Overview

Agent nodes provide access to multiple AI models including:

• GPT-4 - Advanced reasoning and complex tasks
• GPT-3.5 - Fast, efficient for standard tasks
• Claude - Anthropic's AI assistant
• Local Models - Self-hosted LLMs

Configuration

Model Selection

Choose the appropriate model for your use case:

Core Parameters

Prompt

The instruction or question for the AI model:

Analyze the customer feedback and identify:
1. Main concerns
2. Satisfaction level (1-10)
3. Suggested improvements

Data: {{input}}

System Message (Optional)

Sets the AI's role and behavior:

You are a customer service expert. Be concise, 
professional, and focus on actionable insights.

Temperature (0-2)

Controls creativity vs consistency:

• 0: Deterministic, same output each time
• 0.7: Balanced creativity (default)
• 1.5: Creative, varied responses
• 2: Maximum creativity, may be unpredictable

Max Tokens

Maximum response length (1 token ≈ 4 characters):

• Short responses: 50-200
• Standard: 500-1000
• Long form: 2000-4000

Input Data

Agent nodes receive data from previous nodes through {{input}}:

Direct Reference

Summarize this article: {{input.article}}
Customer name: {{input.customer.name}}

JSON Processing

Process this order data:
{{input}}

Extract key information and format as a shipping label.

Prompt Engineering

Effective Prompts

Be Specific

❌ "Process this data"

✅ "Extract customer names, emails, and purchase amounts 
from the JSON data. Format as a CSV with headers."

Provide Structure

Analyze the sales data and provide:

## Summary
[2-3 sentences overview]

## Key Metrics
- Total revenue:
- Number of transactions:
- Average order value:

## Recommendations
1. [First recommendation]
2. [Second recommendation]

Include Examples

Convert product descriptions to marketing copy.

Example:
Input: "Blue cotton t-shirt, machine washable"
Output: "Premium comfort meets easy care in our classic blue cotton tee"

Now convert: {{input.description}}

Common Use Cases

Data Extraction

Extract the following from the email:
- Sender name
- Main request
- Urgency level (low/medium/high)
- Required action

Email content: {{input.emailBody}}

Content Generation

Write a professional response to this customer complaint:
{{input.complaint}}

Tone: Empathetic and solution-focused
Include: Apology, explanation, resolution offer
Length: 100-150 words

Analysis & Classification

Classify this support ticket:

Category: [Technical/Billing/Feature Request/Other]
Priority: [P1-Critical/P2-High/P3-Medium/P4-Low]
Sentiment: [Positive/Neutral/Negative]
Estimated Resolution Time: [hours]

Ticket: {{input.ticket}}

Translation & Transformation

Convert this technical documentation to user-friendly FAQ:

Technical: {{input.documentation}}

Format:
Q: [Question in simple terms]
A: [Clear, concise answer without jargon]

Advanced Techniques

Chain of Thought

Solve this step by step:

1. First, identify the problem type
2. List relevant information
3. Apply appropriate method
4. Verify the solution
5. Provide final answer

Problem: {{input.problem}}

Few-Shot Learning

Learn from these examples then process the new item:

Example 1:
Input: {sku: "ABC123", qty: 5}
Output: "Stock Check Required: ABC123 (5 units)"

Example 2:
Input: {sku: "XYZ789", qty: 0}
Output: "Out of Stock Alert: XYZ789"

New item: {{input}}

Role-Based Responses

System: You are a senior data analyst with 10 years experience.

Prompt: Review this quarterly report and provide executive insights:
{{input.report}}

Focus on: Trends, risks, opportunities, action items

Output Handling

Structured Output

Request specific formats:

Return your analysis as JSON:
{
  "summary": "...",
  "score": 0-100,
  "recommendations": ["...", "..."],
  "nextSteps": "..."
}

Parsing Responses

The Agent node output can be used in downstream nodes:

// In a Code node after Agent
const analysis = JSON.parse(input);
const score = analysis.score;
const urgent = score < 50;

Model-Specific Features

GPT-4 Vision

Analyze images (when available):

Describe the contents of this image.
Identify any text, objects, and overall scene.

Claude's Long Context

Process large documents:

Review this 50-page contract and highlight:
1. Key terms and conditions
2. Potential risks
3. Unusual clauses

Function Calling

Some models support function definitions:

{
  "function": "search_database",
  "description": "Search customer records",
  "parameters": {
    "query": "string",
    "filters": "object"
  }
}

Best Practices

1. Optimize Token Usage

• Be concise in prompts
• Request specific output lengths
• Remove unnecessary context

2. Handle Errors Gracefully

• Add fallback prompts
• Validate AI responses
• Include retry logic for failures

3. Maintain Consistency

• Use system messages for consistent behavior
• Keep temperature low for predictable outputs
• Create prompt templates for repeated tasks

4. Security Considerations

• Never include sensitive data in prompts
• Validate and sanitize AI outputs
• Be cautious with executable code generation

Troubleshooting

"Model timeout"

• Reduce max tokens
• Simplify prompt
• Break into smaller tasks

"Inconsistent outputs"

• Lower temperature setting
• Add more specific instructions
• Include output examples

"Exceeds token limit"

• Shorten input data
• Summarize before processing
• Use a model with larger context

"Poor quality responses"

• Improve prompt clarity
• Add examples
• Try different model
• Adjust temperature

Cost Optimization

Strategies

1. Use appropriate models - GPT-3.5 for simple tasks
2. Cache responses - Reuse for identical inputs
3. Batch processing - Combine multiple requests
4. Prompt optimization - Shorter prompts = lower cost
5. Local models - For high-volume, non-critical tasks

Examples

Customer Service Bot

System: You are a helpful customer service representative.

Prompt: Respond to this customer inquiry:
{{input.inquiry}}

Guidelines:
- Be friendly and professional
- Provide specific solutions
- Include relevant links/resources
- Keep response under 150 words

Content Moderator

Analyze this user comment for:
1. Inappropriate content (yes/no)
2. Sentiment (positive/neutral/negative)
3. Action required (approve/flag/reject)

Comment: {{input.comment}}

Return as: {"appropriate": boolean, "sentiment": string, "action": string}

Data Transformer

Convert this unstructured text into structured data:

Text: {{input.text}}

Extract:
- Names (people and companies)
- Dates and times
- Monetary amounts
- Locations
- Key actions/events

Format as JSON with appropriate keys.

Integration Tips

With Code Nodes

// Process Agent output
const aiResponse = input;
const structured = aiResponse.split('\n').map(line => ({
  item: line.trim(),
  processed: true
}));
return structured;

With Condition Nodes

Use Agent for intelligent routing:

Classify urgency as HIGH, MEDIUM, or LOW:
{{input.message}}

Then branch based on the classification.

With Loop Nodes

Process arrays item by item:

Summarize this review in one sentence:
{{input.currentItem.review}}

Related Topics

• Code Node - Process Agent outputs
• Agents Overview - How agents fit into workflows
• Local AI Models - Run models on your own hardware
• Usage Metering - Track AI usage