Skip to main content

​
Configuring Agents

Fine-tune your agent’s behavior with advanced settings.

​
Core Configuration Options

​
System Prompt

The most important configuration—defines your agent’s behavior, personality, and capabilities. Structure: 
[Who the agent is]
You are a [role] for [company/purpose].

[What the agent does]
Your responsibilities include:
- [Task 1]
- [Task 2]
- [Task 3]

[How the agent should behave]
Your style:
- [Personality trait 1]
- [Personality trait 2]
- [Constraint or guideline]

[What the agent knows]
You have access to:
- [Knowledge source 1]
- [Knowledge source 2]
Example: 
You are a senior customer success manager for TechCo SaaS.

Your responsibilities include:
- Answering product questions
- Troubleshooting technical issues
- Guiding users to documentation
- Escalating billing questions to finance team

Your style:
- Professional yet friendly
- Patient with non-technical users
- Clear, concise explanations
- Always follow up with "Is there anything else I can help with?"

You have access to:
- Product documentation
- Common troubleshooting guides
- Feature comparison charts

​
Temperature

Controls response randomness (creativity vs consistency). Scale: 0.0 to 1.0
  • 0.0-0.2: Deterministic, focused (code, data extraction)
  • 0.3-0.5: Balanced, reliable (most use cases)
  • 0.6-0.8: Creative, varied (content writing)
  • 0.9-1.0: Highly creative, unpredictable (brainstorming)

​
Max Tokens

Maximum length of responses.
  • 500-1000: Short, concise answers
  • 1000-2000: Standard responses (default)
  • 2000-4000: Detailed explanations
  • 4000+: Long-form content
Lower max tokens = faster responses and lower costs

​
Context Window

How much conversation history the agent remembers.
  • Small (4K): Fast, cheap, recent context only
  • Medium (16K): Balanced for most conversations
  • Large (128K-200K): Entire conversation history

​
Advanced Options

Text that signals the agent to stop generating.Example: ---END--- to end responses cleanly
Reduces repetition (0.0-2.0).
  • 0.0: No penalty
  • 1.0: Moderate reduction
  • 2.0: Strong reduction
Encourages topic diversity (0.0-2.0).Higher values = more topic variety
Alternative to temperature for randomness.
  • 0.1: Very focused
  • 0.9: More diverse (default)

​
Configuration Recipes

​
Technical Support Agent

Model: GPT-3.5 Turbo
Temperature: 0.3
Max Tokens: 1500
Context: 16K
Frequency Penalty: 0.5
Presence Penalty: 0.3

​
Creative Content Writer

Model: Claude 3 Opus
Temperature: 0.8
Max Tokens: 3000
Context: 128K
Frequency Penalty: 0.2
Presence Penalty: 0.6

​
Code Assistant

Model: GPT-4
Temperature: 0.2
Max Tokens: 2000
Context: 128K
Frequency Penalty: 0.0
Presence Penalty: 0.0

​
Research Analyst

Model: Claude 3 Sonnet
Temperature: 0.4
Max Tokens: 4000
Context: 200K
Frequency Penalty: 0.3
Presence Penalty: 0.4

​
Knowledge Base

Add information your agent can reference:
  • Upload Files
  • Connect Data
  • PDFs, DOCX, TXT, MD
  • Product documentation
  • Policy guides
  • FAQ documents
  • Max 50MB total

​
Testing Your Configuration

  1. Test with varied inputs - Try different question types
  2. Check edge cases - What happens with unusual inputs?
  3. Verify tone - Does it match your brand?
  4. Review accuracy - Are responses factually correct?
  5. Monitor costs - Track token usage
Always test configuration changes before deploying to production!

​
Configuration Best Practices

✅ Do:
  • Start with defaults and adjust incrementally
  • Document why you chose specific settings
  • Test thoroughly before deploying
  • Monitor performance and iterate
❌ Don’t:
  • Set temperature to extremes without testing
  • Ignore token costs
  • Skip testing edge cases
  • Change too many settings at once

Next: Test Your Agent

Learn how to properly test your configured agent
⌘I