Building Agents

Agent Builder Overview

Access the Agent Builder from the left sidebar to create and manage your agents.

The Agent Builder provides:

• Your Agents: Personal agents you’ve created
• Shared Agents: Organization-wide agents
• Create New Agent: Button to start building
• Search: Find agents by name or description
• Agent Management: Edit, duplicate, or delete agents

Creating Your First Agent

Starting the Creation Process

1. Click “Create New Agent” in the top-right corner
2. You’ll be directed to the agent configuration form
3. The form is organized into sections for different aspects of your agent

Basic Information

The first section covers fundamental agent details.

Agent Name

Purpose: A clear, descriptive name for your agent

Best Practices:

• Use descriptive names that indicate the agent’s purpose
• Include domain or function: “FinServ Compliance Sentry”, “Code Review Assistant”
• Keep it concise but meaningful: 2-5 words ideal
• Consider naming conventions for your organization

Description

Purpose: Explain what your agent does and when to use it

What to Include:

• Primary function: What the agent specializes in
• Use cases: When to use this agent
• Capabilities: What it can help with
• Limitations: What it cannot do

Model & Instructions

This section defines your agent’s intelligence and behavior.

Base Model Selection

Purpose: Choose the underlying large language model

Available Models (vary by deployment):

• gemma-3-27b-it: Google’s instruction-tuned open-weight model built for balanced performance across diverse tasks.
• gpt-oss-120b: OpenAI’s open-weight model designed for powerful reasoning, agentic tasks, and versatile developer use cases.
• Hermes-3-Llama-3.1-8B: General use model that excels at reasoning and multi-turn conversations, with an improved focus on longer context lengths.
• Qwen2.5-Coder-14B-Instruct: Designed to enhance code generation, reasoning, and fixing, making it a powerful tool for developers.
• Qwen2.5-VL-7B-Instruct: Multimodal vision-language mode that excels in analyzing images, video and text.

Choosing the Right Model:

System Prompt

Purpose: Define your agent’s personality, expertise, and behavior

The Most Important Configuration: Your system prompt determines how your agent thinks and responds.

System Prompt Structure

A well-crafted system prompt includes:

1. Role Definition

You are a financial services compliance expert specializing in
regulatory requirements for banking and fintech companies.

2. Core Responsibilities

Your primary responsibilities:
- Answer questions about financial regulations (SOX, GDPR, PCI DSS)
- Assess compliance risks in business processes
- Provide guidance on regulatory requirements
- Suggest compliance best practices

3. Knowledge Scope

You have deep expertise in:
- US financial regulations (Dodd-Frank, SOX, BSA/AML)
- European regulations (GDPR, MiFID II, PSD2)
- Payment card industry standards (PCI DSS)
- Data protection and privacy laws

4. Behavioral Guidelines

When responding:
- Always cite specific regulations when applicable
- Highlight potential risks clearly
- Provide actionable recommendations
- Acknowledge when unsure and suggest consulting legal counsel
- Use clear, professional language

5. Limitations

Important limitations:
- You are not a lawyer; responses are informational, not legal advice
- Always recommend consulting qualified legal professionals for final decisions
- You cannot guarantee regulatory compliance

System Prompt Best Practices

Be Specific: ❌ “You are a helpful assistant” ✅ “You are a cybersecurity expert specializing in OWASP Top 10 vulnerabilities and secure coding practices for web applications”

Define Expertise Boundaries:

You specialize in Python and JavaScript security.
For questions about other languages, acknowledge your
limitations and suggest consulting language-specific experts.

Set Response Format:

Structure your responses as:
1. Summary (2-3 sentences)
2. Detailed Analysis
3. Recommendations (numbered list)
4. References (if using knowledge base)

Include Safety Guidelines:

Never provide information that could:
- Bypass security controls
- Violate privacy regulations
- Harm users or systems
- Violate ethical guidelines

Use Examples in Prompts:

When explaining code vulnerabilities, provide:
- Example vulnerable code
- Explanation of the risk
- Corrected secure code
- Prevention strategies

Streaming Mode

Purpose: Control how responses are delivered

Options:

• Streaming: Responses appear word-by-word in real-time
• Non-Streaming: Complete response delivered at once

When to Use Streaming:

• ✅ Interactive conversations
• ✅ Long-form responses
• ✅ Better perceived performance
• ✅ User can see progress

When to Use Non-Streaming:

• ✅ Short, quick responses
• ✅ API integrations
• ✅ Batch processing
• ✅ When complete response needed before action
• ✅ When using MCP tool calls

Generation Settings (Advanced Settings)

Fine-tune how your agent generates responses.

Temperature

Range: 0.0 to 1.0

Purpose: Controls randomness and creativity in responses

Scale:

0.0 ←────────────────────→ 1.0
More Deterministic    More Creative

Recommended Settings:

Top P (Nucleus Sampling)

Range: 0.0 to 1.0

Purpose: Controls diversity by limiting token choices to a cumulative probability threshold

Scale:

0.0 ←────────────────────→ 1.0
More Focused         More Diverse

How It Works:

• Model ranks all possible next tokens by probability
• Top P selects from tokens that cumulatively reach the threshold
• Lower values = Consider only most likely tokens (focused)
• Higher values = Consider broader range of tokens (diverse)

Recommended Settings:

• 0.9-0.95: Good balance for most applications (default)
• 0.5-0.7: More focused, consistent responses
• 0.95-1.0: Maximum diversity for creative tasks

Top K

Range: 1 to unlimited (typically 1-100)

Purpose: Limit sampling to the K most likely tokens

How It Works:

• Model considers only the K most probable next tokens
• Lower K = More focused, repetitive
• Higher K = More variety, potential for surprises

Recommended Settings:

• 1: Always pick most likely token (deterministic)
• 10-20: Focused, consistent responses
• 40-50: Balanced variety
• 80-100: High diversity

Max Output Tokens

Range: Varies by model (typically 100-4000+)

Purpose: Maximum number of tokens the model can generate in a single response

Understanding Tokens:

• Tokens ≠ Words
• Approximate: 1 token ≈ 0.75 words (English)
• Example: “Hello, world!” ≈ 3-4 tokens

Recommended Settings:

Resources & Knowledge

Configure how your agent accesses and uses organizational knowledge.

Allow File Uploads

Purpose: Enable users to upload files during conversations

Toggle: On/Off

When Enabled:

• Users can upload documents while chatting
• Agent can read and analyze file contents
• Files persist throughout the conversation

Knowledge Bases

Purpose: Attach organizational knowledge to your agent

Dropdown: Select one or more knowledge bases

How It Works:

1. User asks a question
2. System performs semantic search on attached knowledge bases
3. Relevant documents/chunks retrieved
4. Context added to agent’s prompt
5. Agent generates response using knowledge

Customize Search Settings

Toggle: On/Off

When enabled, you can customize:

Minimum Relevance Score

Range: Off to Stricter

Purpose: Only include results above this similarity score

Settings:

• Off: Return all results, no filtering
• Low (0.3-0.5): Include loosely related documents
• Medium (0.5-0.7): Balance precision and recall
• High (0.7-0.85): Only highly relevant documents
• Strict (0.85-1.0): Only near-exact matches

Maximum Results

Range: 1 to 20

Purpose: Maximum number of document chunks to include in each search

Recommended Settings:

• 1-3: Focused, specific queries
• 5-10: Standard (balanced)
• 10-20: Complex queries needing broad context

Maximum Context Size

Range: Typically 1,000 to 50,000+ characters

Purpose: Maximum characters from knowledge base results to include in agent context

Example: 18000 characters ≈ 4,500 tokens ≈ 3,000 words

Access Control & Security

Configure who can use your agent and what security features are enabled.

Share With Organization

Toggle: On/Off

Purpose: Make this agent visible to all users in your organization

When to Share:

• ✅ General-purpose assistants
• ✅ Company-wide knowledge agents
• ✅ Standard operating procedure guides

Prompt Injection Protection

Toggle: On/Off

Purpose: Detect and block prompt injection attacks in user messages

What It Protects Against:

• Users trying to override system prompt
• Attempts to extract system instructions
• Malicious prompt manipulation
• Jailbreak attempts

PII Detection & Replacement

Toggle: On/Off

Purpose: Automatically detect and replace personally identifiable information in messages

What It Detects:

• Names, email addresses, phone numbers
• Social Security Numbers
• Credit card numbers
• Addresses and IP addresses
• Dates of birth

Topic Filter

Control what subjects your agent can discuss.

Purpose

Topic filters ensure your agent stays focused on its intended purpose and doesn’t respond to off-topic or inappropriate queries.

Configuration Options

Allowed Topics (Whitelist approach):

• Define specific topics agent should address
• Agent politely declines off-topic requests

Blocked Topics (Blacklist approach):

• Define topics agent should refuse
• Agent declines these even if otherwise relevant

Saving and Testing Your Agent

Saving Your Agent

1. Review all settings one final time
2. Click “Create Agent” or “Save Changes”
3. Agent is now available for use
4. Appears in “Your Agents” section

Testing Your Agent

After creation, thoroughly test your agent:

Basic Functionality:

1. Start a new chat with your agent
2. Ask typical questions from your use case
3. Verify responses are appropriate
4. Check tone and style match expectations

Next Steps

Now that you can build agents:

1. Configure MCP - Connect your agent to external tools and live data sources
2. Create Knowledge Bases - Add organizational knowledge to your agents
3. Share with Team - Enable organization sharing and collaborate