Creating a Custom Agent - Julius

Creating Your First Agent

How Custom Agents Work

When you configure a custom agent, you’re building a specialized assistant that works within Julius’s context window. Here’s how all the pieces fit together: How Custom Agents Work Your Agent Configuration Lives in the System Prompt: Your custom agent’s configuration—including Instructions, Training Tasks, and a curated version of your Knowledge Base—becomes part of Julius’s system prompt. This means every response is guided by your specific setup. The Knowledge Base Curator: Your full knowledge base (up to 10,000 characters) stays separate. When you ask a question, a specialized Knowledge Base Curator Sub-Agent analyzes your message against your complete knowledge base and extracts only the relevant information. This curated subset is what gets added to the system prompt, keeping responses focused and efficient. The Complete Picture: Your agent combines:

System Prompt: Your instructions, training tasks, and curated knowledge
Chat History: Your ongoing conversation context
All of Julius’s standard capabilities (code execution, visualizations, data analysis, etc.)

This means your agent always has the right context for consistent, relevant responses without being overwhelmed by unnecessary information.

Configuration Options

Agent Configuration

Agent Name & Description

Agent Name: Choose a clear, descriptive name for your agent. Description: Write a brief description that explains what your agent does. This helps you and your team quickly identify the right agent for the task.

Behavior Settings

Control how your agent responds with the behavior slider:

Short: Concise, to-the-point responses
Standard (Default): Balanced responses with appropriate detail
Detailed: Comprehensive responses with thorough explanations

Output Format

Choose what type of output your agent prioritizes:

All (default): Balanced approach to all output types
Visualizations First: Prioritizes creating charts and graphs
Files/DataFrames First: Focuses on data tables and exports
Insights First: Emphasizes written analysis and takeaways

Instructions

Character Limit: 1,000 characters Instructions define your agent’s general methodology and approach to analysis. Think of this as your agent’s working philosophy that applies to every interaction. Example:

Always start by asking clarifying questions about the data. 
Focus on actionable insights for marketing campaigns. 
Highlight trends and anomalies in customer behavior. 
Create visualizations before providing detailed explanations.

Training Tasks & Approach

Training Tasks Training tasks teach your agent how to handle specific questions with concrete, step-by-step examples. When you ask a similar question later, your agent recognizes the pattern and applies what you taught it. You can add up to 5 training tasks. Each task has two parts:

Task: The question or request (e.g., “How many active subscriptions do we have?”)
How to approach: Step-by-step instructions for completing this task

Example: Task: “Show me monthly revenue trends” How to approach:

Query the transactions table, group by month using DATE_TRUNC.
Calculate total revenue per month.
Create a line chart showing the trend.
Highlight any months with >10% change from previous month.
Provide a brief summary of the overall trend.

Why Use Training Tasks?

Ensures consistency on repeated tasks
Teaches best practices to your agent
Reduces the need for detailed instructions each time
Great for onboarding team members to your analytical standards

Knowledge Base

Character Limit: 10,000 characters The knowledge base holds all the context about your company, data, and domain. You don’t need to reference it manually. When you ask a question, Julius reads through your knowledge base and pulls only the relevant information for that specific question. What to Include:

Company background and business model
Data structure and key tables/fields
Industry-specific terminology
Important metrics and how they’re calculated
Common data issues or quirks
Relationships between different data sources

Example:

Our company is a B2B SaaS platform selling project management software.

Pricing Tiers:
- Starter: $49/month (up to 10 users)
- Professional: $149/month (up to 50 users)  
- Enterprise: Custom pricing (unlimited users)

Key Metrics:
- MRR (Monthly Recurring Revenue): Sum of all active subscription values
- Churn Rate: Cancelled subscriptions / total active subscriptions
- CAC (Customer Acquisition Cost): Marketing spend / new customers

Database Structure:
- users table: Contains customer account information
- subscriptions table: Active and cancelled subscription data
- transactions table: Payment history and invoices
- usage_logs table: Product usage and feature adoption data

Important Notes:
- Free trial users have subscription_status = 'trial'
- Enterprise deals are marked as custom_pricing = true
- Usage data is logged daily at midnight UTC

Web Search

Toggle whether your agent can use web search to find current information beyond its training data. Useful for agents that need to reference current events, prices, or frequently changing information.

Saving Your Agent

Once you’ve configured your agent:

Click Save to store your configuration
Click Use Agent to activate it immediately and start a new chat
Or click Share with Team to make it available to team members

Best Practices

Start Simple

Begin with basic instructions and a few training tasks. Add complexity as you understand what works.

Be Specific

Vague instructions lead to inconsistent results. Clear, specific guidance helps your agent deliver exactly what you need.

Use Real Examples

Base training tasks on actual questions you regularly ask. This ensures your agent handles real-world scenarios.

Iterate and Improve

Monitor your agent’s performance and refine the configuration. Add new training tasks as you discover common patterns.

Document Your Data

A well-maintained knowledge base makes a huge difference. Keep it updated as your data structure changes.