C3 AI Documentation Home

Create Agents from Scratch

Use this tutorial to create a new agent from scratch in the C3 Agentic AI Platform. This tutorial provides complete control over your agent's behavior, capabilities, and integration with C3 AI applications.

You will complete these tasks in sequence:

  1. Create a new agent - Establish the foundational structure.
  2. Configure agent settings - Define behavior and capabilities.
  3. Add comprehensive documentation - Provide extended knowledge and guardrails.
  4. Test your agent configuration - Validate behavior against requirements.
  5. Deploy your agent - Make it available for organizational or end-user access.

Each step serves a specific purpose in creating an agent that performs reliably in production environments while maintaining consistency with your organization's standards and expectations.

Create a new agent

Create a new agent to set up the basic information that determines how your agent works and how others in your organization will find and use it.

The Agent Gallery displays three sections:

  • Drafts: Agents in development.
  • Store: Ready-to-use agents.
  • Templates: Pre-built starting points.

Understanding these sections helps you assess whether your requirements might be met by existing solutions before investing time in custom development. For details, see Gallery.

Start building your agent by creating a new draft:

  1. Select the blue Create Agent button in the top right corner of the Gallery page.

  2. In the Create Agent modal, provide the following information with careful consideration:

    Agent details:

    • Name: Enter a clear and descriptive name that communicates purpose and scope.

      Choose names that are specific (for example, "Financial Report Analyzer" not "Helper Agent"), professional, memorable, and scalable for related agents.

    • Description: Write a comprehensive explanation of what your agent will do, including specific use cases and limitations.

      Include primary use cases, key capabilities, intended audience, and scenarios where the agent should NOT be used.

      Example: "Analyzes quarterly financial reports to identify trends, anomalies, and key performance indicators. Designed for finance teams and executives who need rapid insights from structured financial data. Does not provide investment advice or handle sensitive customer financial information."

    • Category: Select an existing category or create a new one that best fits your agent's purpose. Categories help organize agents in the Agent Gallery and make them easier to find.

  3. Select Create to generate your new agent draft.

Your agent now appears in the Drafts section where you can safely experiment with configurations without affecting other users or production systems.

Configure agent settings

The Agent Workbench opens after you create your agent. The workbench displays three panels:

  • Configuration panel (left) - Contains all behavioral controls for your agent including objectives, prompts, model selection, memory settings, and tools.
  • Chat panel (center) - Provides a live chat interface to test your agent's responses in real time.
  • Trace panel (right) - Displays detailed execution information for debugging and monitoring. This panel is collapsed by default.

All panels are collapsible. The Trace panel is collapsed by default.

For complete interface details, see Agent Workbench.

Define agent objective and prompts

The objective and prompts section establishes your agent's core behavioral programming, including personality, communication style, and operational boundaries. These settings directly influence every interaction your agent has with users.

Configure your agent's behavior:

  1. Navigate to the agent workbench and locate the configuration panel on the left side of the screen.

  2. Write clear, specific instructions in the Objective & Prompts section that establish your agent's role, responsibilities, and behavioral guidelines. For a customer support agent, consider this example:

    Text
    You are a professional customer support agent for our technology company. Your primary responsibility is providing accurate, helpful assistance to customers experiencing technical issues or seeking product information. Maintain a friendly but professional tone in all interactions. When you can resolve issues directly using available information, provide step-by-step guidance. When issues require escalation or access to customer-specific data, clearly explain next steps and how to reach appropriate human support. Always acknowledge customer concerns and demonstrate empathy while working toward solutions. Never make promises about features, timelines, or policies without explicit documentation to support your statements.

  3. Test your prompt by observing its effects on agent responses in the chat panel.

Select language model

Select your preferred language model in the Model section. Available models may include:

  • Claude 4.5 Sonnet
  • Gemini 2.5 Pro
  • GPT 5

Available models vary based on your application configuration and cluster setup. Contact your administrator if you need access to specific models. For information on configuring LLM clients, see Configure LLM Clients.

Language model selection affects your agent's capabilities, response quality, and operational costs.

Configure memory settings

Memory configuration determines how your agent maintains context across conversations, affecting user experience and interaction quality.

Set up conversation memory using the Short Term Memory toggle:

  • Memory enabled: The agent references earlier messages and maintains conversation flow, enabling natural, multi-turn conversations. Use for customer support with complex issues, educational agents guiding multi-step processes, or creative tasks building upon previous work.

  • Memory disabled: The agent processes each message independently as isolated interactions. Use for high-security environments requiring conversation isolation, simple question-answer scenarios, or compliance scenarios requiring interaction independence.

For most applications, enable memory to provide superior user experience.

Attach tools and integrations

Tools transform your agent from a conversational interface into an active system participant capable of performing real work. Tool selection should align with your agent's intended purpose and existing infrastructure.

Select the plus icon (+) in the Tools section to add tools. Tools extend your agent with capabilities like data access, external API integrations, and specialized computations. For creating custom python tools, see Create Python Tools from the UI.

Each tool addition should solve a specific user problem or enable a capability that justifies the additional complexity.

Add agent Documentation

Agent documentation describes what your agent does and its capabilities for other users. When someone opens your agent in the workbench, they can read this documentation to understand the agent's purpose, capabilities, and how to use it effectively.

Documentation is also available from the gallery page when you hover over the "i" icon on the agent card.

Document your agent so other users understand what it does and how to use it:

  1. Select the Documentation button in the upper right corner of the workbench. This opens a text editor where you can describe your agent for other users. You can also access and edit this documentation from the Agent Gallery by hovering over the i icon on your agent card and selecting the edit option.

  2. Select the pencil icon to enter edit mode.

  3. Describe what your agent does. Explain the agent's primary purpose, the types of questions it can answer, and the tasks it can perform.

  4. Provide example interactions that show users how to interact with the agent and what kind of responses to expect.

  5. Define limitations and boundaries. Explain what topics the agent cannot help with, what actions it cannot perform, and when users should seek alternative resources.

  6. Document permission requirements and access dependencies:

    • What user permissions the agent requires to function fully.
    • Which features depend on specific role-based access controls.
    • How the agent behaves when users lack certain permissions.
    • What data access levels affect agent responses.

    Use this example as a foundation for creating agent documentation:

    Markdown
    Agent Purpose:
This agent provides customer support for our technology products. It answers questions about product features, troubleshoots common technical issues, and guides users through setup processes.

What This Agent Can Do:
- Answer questions about product features and specifications.
- Provide step-by-step troubleshooting guidance for common technical issues.
- Guide users through product setup and configuration.
- Access product documentation and knowledge base articles.
- Create support tickets for issues requiring human assistance.

Available Tools and Integrations:
- Product documentation database.
- Knowledge base search.
- Support ticket creation system.
- Product specification lookup.

Required Permissions:
- Read access to product documentation database.
- Knowledge base search privileges.
- Support ticket creation rights (for ticket generation feature).
- Product catalog access for specification lookups.

Example Interactions:
User: "What features are included in the premium plan?"
Agent: "The premium plan includes features A, B, and C. Would you like detailed information about any specific feature?"

User: "I'm having trouble connecting to the service."
Agent: "I can help troubleshoot that. First, let's check your network connection. Can you confirm you're connected to the internet?"

Limitations:
- Cannot access individual customer account information.
- Cannot process refunds or billing changes.
- Cannot modify user accounts or permissions.
- Cannot provide support for third-party integrations.
- For account-specific issues, users must contact human support.

Permission-Based Limitations:
- Users without documentation access receive limited troubleshooting guidance.
- Ticket creation fails for users without support system permissions.
- Product specification details require catalog access rights.
- Agent responses vary based on user's data access levels.

When to Use This Agent:
- Quick answers about product features.
- Step-by-step troubleshooting assistance.
- Guidance on product setup and configuration.
- Finding relevant documentation.

  7. Select Save to preserve your changes and integrate the documentation into your agent's knowledge base.

Test your agent configuration

Testing validates that your configuration choices produce the intended agent behavior and identifies areas requiring refinement before deployment.

Use the chat panel

The chat panel in the center of the workbench provides immediate feedback on your configuration changes through real-time conversation simulation.

Execute a comprehensive testing sequence:

  1. Self-awareness verification:

    • "What can you help me with?" - Confirms your agent accurately describes its capabilities and limitations.

  2. Tool integration validation:

    • "What tools do you have access to?" - Verifies tool attachments and integration functionality.

  3. Domain expertise assessment:

    • Questions specific to your agent's intended purpose that require domain knowledge.

  4. Boundary testing:

    • Requests clearly outside your agent's intended purpose.
    • Questions requiring information not provided in documentation.
    • Scenarios where escalation to human support is appropriate.

  5. Edge case testing:

    • Unclear or incomplete questions.
    • Invalid tool inputs or connection failures.
    • Contradictory user instructions.

Review execution traces

The Trace panel on the right side of the workbench displays detailed execution information for each interaction. Expand the panel to view operation sequences, execution timing, and data exchanged during agent processing. For trace interpretation guidance, see Debug using Traces and Spans.

Deploy your agent

Deployment transforms your tested agent configuration into an operational system that can serve real users. You can deploy from the Agent Workbench or from the Store.

Choose deployment path

After completing agent configuration, select Finish Configuration in the agent workbench. You have two deployment options:

  • Store - Makes your agent available for others in your organization to discover, evaluate, and customize for their specific needs. The agent gets saved but remains inactive. Choose this option to share your agent without deploying it live.

  • Store & Deployment - Makes your agent available to your organization and deploys it live for end users. Real users can interact with it immediately. Choose this option when your agent is ready for production use.

Move to Store

To share your agent with your organization without making it live:

  1. Select Move to Store in the Finish Configuration modal.

  2. Complete the store information form:

    • Name: Enter a descriptive name for your agent
    • Description: Provide details about the agent's capabilities and use cases

  3. Select Move to Store to save the agent.

Deploy to production

You can deploy an agent in two ways:

Deploy directly from Agent Workbench

To deploy immediately after configuration:

  1. Select Store & Deployment in the Finish Configuration modal.

  2. Complete the store information form, then select Deploy Configuration.

  3. The Deployment Configuration modal appears. Complete the following fields:

    • Agent name: Pre-filled with the agent's name and not editable
    • Deployment type: Choose New Deployment to create a new deployment or Update existing deployment to redeploy to the same endpoint
    • Deployment name: Enter a unique identifier for this deployment (only appears when you select New Deployment)
    • Node pool: Select compute resources (contact your administrator for options)
    • Thread pool size: Set concurrent conversation capacity based on your expected workload
    • Tracing mode: Enable or disable execution logging
    • Publish to package: Select this checkbox to make the deployment available across applications that use your package. You must be in development mode to do this.

    Selecting Update existing deployment redirects all existing applications and users to the new agent version immediately. This affects production applications. Use this option with caution.

  4. Select Deploy to start the deployment process.

Deploy from Store

To deploy an agent that is already in the Store:

  1. Navigate to Agents > Gallery.
  2. Locate your agent in the Store section.
  3. Select Deploy next to your agent.
  4. Configure the deployment settings as listed above.
  5. Select Deploy.

The system provisions resources and activates your agent within 2-3 minutes.

Verify deployment

After deployment:

  1. Navigate to Agents > Deployments.
  2. Confirm status shows "Running".
  3. Select Chat to test basic functionality.

Learn more about Deployments in Agent Deployment Management

Production deployments are immediately accessible to end users with access to your application. Test thoroughly in the workbench before deploying business-critical agents.

See also

Was this page helpful?