​

Building Agents

Learn how to create, configure, and publish custom agents for your specific needs.

​

Getting Started

​

Prerequisites

Before building an agent, ensure you have:

• At least one LLM configured - Required for agent to function
• Optional: Knowledge bases, skills, or custom tools
• Optional: Image generation models for visual tasks

​

Creating an Agent

​

Step 1: Navigate to Build Page

Go to Build in the left sidebar to access the agent management interface.

​

Step 2: Create New Agent

Click New Agent to open the agent configuration editor. The editor has a two-column layout:

• Left panel: Configuration options
• Right panel: Live preview chat interface

​

Agent Configuration

​

Basic Information

Name (Required)

• Maximum 200 characters
• Must be unique within your account
• Choose a descriptive name for your agent

Description (Optional)

• Brief explanation of what the agent does
• Helps users understand the agent’s purpose

Logo (Optional)

• Upload an image to represent your agent
• Displayed in agent listings and chat interface

​

Execution Mode

Choose how your agent processes tasks: React Mode (Default)

• ReAct loop: Reasoning → Acting cycle
• Best for: Most general-purpose tasks
• Agent thinks, takes action, observes results, repeats

Graph Mode

• DAG planning-execution
• Best for: Complex, multi-step workflows
• Agent plans all steps first, then executes

​

AI Models

Configure models for different purposes: General Model (Required)

• The main LLM for all task execution
• Handles planning, reasoning, and decision making
• Example: Claude 4.6 Sonnet, GPT 5.2

Fast Model (Optional)

• Lightweight model for simple operations
• Reduces cost and improves speed
• Automatically used for simple steps identified during planning

Visual Model (Optional)

• Multimodal LLM for image understanding
• Required for image-related tasks
• Enables tools like understand_images, describe_images

Compact Model (Optional)

• Specialized model for compressing long conversations
• Maintains context across extended sessions
• Triggered when conversation exceeds token threshold

​

Capabilities

Knowledge Bases (Optional)

• Add domain-specific knowledge to your agent
• Agent can search and retrieve relevant information
• Upload documents or select existing knowledge bases

Skills (Optional)

• Pre-built capabilities for specific tasks
• Enhance agent with specialized functionality
• Examples: Web search, database operations

Tools (Optional)

• Select tool categories to enable:
  • basic: Basic tools (web search, calculator, etc.)
  • file: File upload and management
  • vision: Image understanding
  • image: Image generation and editing
  • knowledge: Knowledge base operations
  • mcp: MCP protocol tools
  • browser: Browser automation
  • ppt: PowerPoint processing
  • office: Office document processing

​

Suggested Prompts

Provide example prompts to help users get started:

• “Analyze this sales data and create a report”
• “Generate a marketing plan for our new product”
• “Review these documents and summarize key points”

These prompts appear as suggestions when users chat with your agent.

​

Testing Your Agent

​

Live Preview

Use the preview panel on the right to test your agent in real-time:

1. Type a message in the chat input
2. Send and observe how your agent responds
3. Upload files to test file handling
4. Iterate on configuration based on results

The preview uses your current configuration without saving:

• Test different settings before committing
• See immediate results
• Verify tool access and knowledge retrieval

​

What to Test

• Response quality: Does the agent understand the task?
• Tool usage: Are the right tools being used?
• Knowledge retrieval: Does it find relevant information?
• Execution mode: React vs Graph behavior
• Model selection: Are the configured models working?

​

Saving Your Agent

​

Auto-Save

Your configuration is automatically validated but not saved until you explicitly save.

​

Manual Save

Click Save to persist your agent configuration:

• Agent is saved with draft status
• Only you can see and edit draft agents
• Saved agents appear in your Build page list

​

Editing Existing Agents

1. Go to Build page
2. Click Edit on the agent card
3. Make changes to configuration
4. Click Save to update

​

Publishing Your Agent

​

When to Publish

Publish your agent when:

• Configuration is complete and tested
• Ready to share with others
• Production-ready workflows

​

Publishing Process

1. Save your agent first
2. Click Publish button
3. Agent status changes from draft to published
4. Publication timestamp is recorded
5. Agent becomes visible to authorized users

​

Unpublishing

To unpublish an agent:

1. Go to Build page
2. Click Unpublish on the published agent
3. Agent status returns to draft
4. Agent is no longer accessible to users

​

Chatting with Your Agent

​

Starting a Conversation

Once published, users can chat with your agent:

1. Find the agent in the agents list
2. Click on the agent to open its detail page
3. Type a message or use a suggested prompt
4. Agent responds using its configuration

​

Agent Execution

When chatting, the agent:

• Uses its configured models for reasoning
• Follows its execution mode (React or Graph)
• Accesses enabled tools based on tool categories
• Retrieves information from knowledge bases
• Applies its instructions and skills

​

Task Creation

Each conversation creates a task:

• Task is linked to the agent
• Execution uses agent’s configuration
• Results are saved in task history
• Can view full execution trace

​

Agent Templates

​

Using Templates

Jump-start agent creation with pre-built templates:

1. Go to Build page
2. Click New Agent
3. Select a template
4. Template populates configuration
5. Customize as needed

​

Available Templates

Common agent templates include:

• Customer Support: FAQ handling, documentation search
• Data Analyst: Data processing, visualization, reporting
• Research Assistant: Information gathering, synthesis
• Content Creator: Writing, editing, image generation

​

Best Practices

​

Start Simple

• Begin with React mode for most use cases
• Enable only necessary tool categories
• Add knowledge bases after testing basic functionality
• Use preview to iterate quickly

​

Define Clear Purpose

• Specific name and description
• Focused skill set
• Relevant suggested prompts
• Appropriate tool access

​

Test Thoroughly

• Use preview before publishing
• Test with various inputs
• Verify tool access is appropriate
• Check knowledge retrieval accuracy

​

Iterate and Improve

• Monitor agent performance
• Gather user feedback
• Update configuration as needed
• Republish after changes

​

Troubleshooting

​

Agent Not Responding

Check:

• General model is configured
• Model has available API quota
• Network connectivity is working

​

Tools Not Available

Verify:

• Tool categories are selected
• Models for specific tools are configured (e.g., Visual model for vision tools)
• Tool permissions are properly set

​

Knowledge Not Retrieved

Check:

• Knowledge bases are attached
• Documents are indexed
• Embedding model is configured
• Search queries are specific

​

Preview Not Working

Try:

• Refreshing the page
• Re-opening the editor
• Checking browser console for errors
• Verifying model configuration

​

Next Steps

• Templates - Explore pre-built agent templates
• Knowledge Base - Add domain knowledge to your agent
• Tools - Understand available tools
• Models - Configure AI models for your agent
• Tasks - Learn about task execution