Skip to Content
DocsGuidesCreating Agents

Creating Agents

A step-by-step guide to planning, designing, and creating effective AI agents in Huf.

Planning Your Agent

Before creating an agent, plan its purpose, scope, and capabilities.

Define the Role

Ask yourself:

  • What task should this agent perform?
  • What is its primary responsibility?
  • What problems does it solve?
  • Who will use it?

Example:

  • Role: Customer Support Assistant
  • Purpose: Answer customer questions and create support tickets
  • Users: End customers and support team
  • Scope: FAQ, ticket creation, basic information lookup

Define Success Criteria

What does success look like?

  • Response time: Under 30 seconds
  • Accuracy: 95% correct answers
  • User satisfaction: Positive feedback
  • Ticket reduction: 50% fewer manual tickets

Identify Required Capabilities

What can the agent do?

  • Answer questions about products
  • Look up customer information
  • Create support tickets
  • Check order status
  • Provide account information

Choose the Right Model

Consider:

  • Task complexity: Simple tasks need cheaper models
  • Response quality: Complex reasoning needs powerful models
  • Cost: Balance capability with budget
  • Speed: Real-time needs faster models

Guidelines:

  • Simple FAQ: GPT-3.5 Turbo or Claude Haiku
  • Complex analysis: GPT-4 Turbo or Claude Opus
  • High volume: Cheaper models with good instructions
  • Critical decisions: More powerful models

Writing Instructions

Instructions are the most important part of your agent. They define behavior, tone, and capabilities.

Structure

Good instructions include:

  1. Role definition: Who the agent is
  2. Responsibilities: What it does
  3. Capabilities: What tools it has
  4. Guidelines: How to behave
  5. Examples: Sample interactions

Example: Well-Structured Instructions

You are a customer support assistant for an e-commerce platform.

Your responsibilities:
1. Answer questions about products, orders, and policies
2. Look up customer and order information
3. Create support tickets for issues you cannot resolve
4. Provide clear, helpful responses

Available tools:
- get_customer: Look up customer information
- get_order: Retrieve order details
- create_support_ticket: Create tickets for complex issues
- check_product_availability: Check if products are in stock

Guidelines:
- Always be polite and professional
- If you don't know something, say so and create a ticket
- Provide specific information, not vague answers
- Use customer names when addressing them
- Escalate urgent issues immediately

Examples:
User: "What's the status of my order?"
You: "I'll check that for you. What's your order number?"

User: "I need to return an item"
You: "I can help with that. Let me look up your order details..."

Common Mistakes

Avoid:

  • Vague instructions: “Help users” (too broad)
  • Missing context: No mention of available tools
  • Unclear boundaries: Doesn’t know when to escalate
  • No examples: Agent doesn’t understand expected behavior

Instead:

  • Be specific: “Answer questions about orders, products, and shipping”
  • Mention tools: “Use get_order tool to retrieve order details”
  • Set boundaries: “If issue is complex, create a support ticket”
  • Include examples: Show expected interactions

Selecting Tools

Choose tools that match your agent’s responsibilities.

Tool Selection Process

  1. List required capabilities

    • What data does the agent need?
    • What actions should it take?
    • What external systems to integrate?

  2. Map to tools

    • Data retrieval → Get Document, Get List
    • Data creation → Create Document
    • Data updates → Update Document
    • External calls → HTTP Request
    • Complex logic → Custom Tools

  3. Minimize tool count

    • Only assign necessary tools
    • More tools = more tokens = higher cost
    • Focused agents perform better

Example Tool Assignment

Customer Support Agent:

  • get_customer (Get Document)
  • get_order (Get Document)
  • get_list_tickets (Get List)
  • create_support_ticket (Create Document)
  • send_notification (HTTP Request)

Total: 5 tools (focused and sufficient)

Tool Descriptions Matter

Write clear tool descriptions:

  • What the tool does
  • When to use it
  • What parameters it needs
  • What it returns

Example:

Tool: get_customer
Description: Retrieve detailed customer information including 
contact details, account status, and order history. Use this 
when users ask about their account or when you need customer 
context for support.

Configuring Settings

Temperature

Controls randomness:

  • 0.0-0.3: Very focused, deterministic
    • Use for: Data retrieval, validation, structured tasks
  • 0.7: Balanced (default)
    • Use for: General conversations, customer support
  • 1.0-2.0: Creative, varied
    • Use for: Content generation, brainstorming

Recommendation: Start with 0.7, adjust based on results.

Top P

Alternative to temperature:

  • 1.0: Full vocabulary (default)
  • Lower values: More focused word choices
  • Use with temperature: Not recommended (use one or the other)

Conversation Settings

Persist Conversation:

  • Enable for multi-turn interactions
  • Disable for independent requests
  • Affects token costs

Persist per User:

  • Enable for user-specific history
  • Disable for shared conversations
  • Important for triggers

Testing Your Agent

Before Production

Test with sample scenarios:

  1. Common use cases
  2. Edge cases
  3. Error scenarios
  4. Tool usage
  5. Response quality

Testing Checklist

  • Agent understands common questions
  • Tools are called correctly
  • Responses are accurate
  • Error handling works
  • Instructions are followed
  • Tone is appropriate
  • Escalation works when needed

Iterative Improvement

Process:

  1. Create agent with initial configuration
  2. Test with real scenarios
  3. Review Agent Run logs
  4. Refine instructions
  5. Adjust tools if needed
  6. Tune temperature
  7. Test again
  8. Deploy when satisfied

Best Practices

Instructions

Do:

  • Be specific about role and responsibilities
  • Include step-by-step guidance
  • Provide examples of good responses
  • Define clear boundaries
  • Mention available tools
  • Set tone and style expectations

Don’t:

  • Be vague or ambiguous
  • Assume the AI knows your business logic
  • Forget to mention tool usage
  • Mix multiple unrelated roles
  • Use technical jargon users won’t understand

Tool Assignment

Do:

  • Assign only necessary tools
  • Use specific DocTypes per tool
  • Write clear tool descriptions
  • Test tool calls thoroughly

Don’t:

  • Assign too many tools
  • Use generic “get any document” tools
  • Forget to test tool permissions
  • Overlook tool parameter requirements

Model Selection

Do:

  • Match model to task complexity
  • Consider cost vs capability
  • Test with cheaper models first
  • Upgrade if needed

Don’t:

  • Use expensive models for simple tasks
  • Underestimate model requirements
  • Ignore token costs
  • Forget to monitor performance

Common Patterns

FAQ Agent

Pattern:

  • Simple instructions with Q&A examples
  • Get Document tools for data lookup
  • Low temperature (0.3-0.5)
  • No conversation persistence

Use Case: Answer common questions quickly

Support Agent

Pattern:

  • Detailed instructions with escalation rules
  • Multiple tools (lookup, create tickets)
  • Medium temperature (0.7)
  • Conversation persistence enabled

Use Case: Handle support interactions

Analysis Agent

Pattern:

  • Focused on data analysis
  • Get List tools with filters
  • Custom tools for calculations
  • Low temperature (0.3)
  • Scheduled triggers

Use Case: Generate reports and insights

Automation Agent

Pattern:

  • Task-focused instructions
  • Document manipulation tools
  • Very low temperature (0.2)
  • Document event triggers
  • No conversation persistence

Use Case: Automate document processing

Troubleshooting

Agent doesn’t follow instructions:

  • Make instructions more specific
  • Add examples
  • Simplify language
  • Break into smaller steps

Wrong tools being used:

  • Improve tool descriptions
  • Add tool usage guidance to instructions
  • Test tool calls individually
  • Review Agent Run logs

Responses are inconsistent:

  • Lower temperature
  • Add more examples
  • Be more specific in instructions
  • Test with different prompts

High costs:

  • Use cheaper models
  • Reduce tool count
  • Limit conversation history
  • Optimize instructions (shorter)

What’s Next?

Questions? Check GitHub discussions .

Last updated on
Example AgentsManaging Agents