AskCodi Custom Agents

What Are Custom Agents?

Custom agents are AI assistants with specific instructions, tools, and behavior settings.

An agent is like a specialized AI with:

• Pre-loaded instructions (system prompt)
• Configured model settings (temperature, tokens, etc.)
• Available tools (search, context, MCP servers)
• Middleware processing (PII detection, fallbacks)
• Prompt stacks (multi-prompt context)

Instead of configuring these settings every time, you create an agent once and reuse it for specific tasks.

Example Agents:

• Code Reviewer: Reviews code with strict linting rules, low temperature, reasoning nodes
• Creative Writer: High temperature, playful tone, no technical tools
• Data Analyst: Access to search tools, context awareness, structured reasoning
• Support Bot: PII detection middleware, fallback to human agent, specific instructions

Creating an Agent

From the Agents Page

1. Click "Agents" in the left sidebar (or wherever agent management is)
2. Click "Create Agent" button
3. Fill in the comprehensive form

From the Chat Sidebar

1. Click the floating arrow to open the right sidebar
2. Expand the "Agent Builder" section
3. Click "Create Agent" inline
4. Same form as above, but stays in the sidebar
5. Newly created agents appear models dropdown after reload.

Basic Configuration

• Name: Unique identifier (e.g., "code-reviewer-strict")
• Display Name: User-friendly name (e.g., "Code Reviewer")
• Description: What this agent does
• Agent Type: Private or Community
• Provider Model: Select from available models with API keys
• Instructions: The system prompt for this agent
• Show Instructions: Whether to display instructions to users

Advanced Configuration Tabs

The agent form uses tabs to organize complex settings:

Model Overrides Tab

Fine-tune model behavior:

• Temperature (0.0 - 2.0): Randomness in responses
  • 0.0 = Deterministic, focused
  • 1.0 = Balanced creativity
  • 2.0 = Highly creative, unpredictable
• Max Tokens: Maximum response length
• Top P (0.0 - 1.0): Nucleus sampling threshold
• Frequency Penalty (-2.0 - 2.0): Reduce repetition
• Presence Penalty (-2.0 - 2.0): Encourage new topics
• Stop Sequences: Custom stop tokens

All sliders use the specified ranges. Only changed values are saved (not defaults).

Tools Tab

Enable capabilities for the agent:

• Search: Allow web search queries
• Context: Provide conversation context
• MCP (Model Context Protocol): External tool integration
  • Enable MCP toggle
  • Enter MCP API key
  • Configure MCP server endpoint

Middleware Tab

Add processing layers:

• PII Detection: Automatically detect and flag personal information
• Model Fallback: Switch to backup model if primary fails
• Custom Middleware: Add your own processing functions

Each middleware has its own configuration options.

Nodes Tab

Configure reasoning and review nodes:

• Reasoning Node: Multi-step thinking before responding
  • Enable toggle
  • Select reasoning model
• Review Node: Self-review responses before sending
  • Enable toggle
  • Select review model

Nodes add processing steps that improve response quality at the cost of latency.

Metadata Tab

Organizational information:

• Tags: Categorize agents for search
• Version: Track agent iterations
• API Calls Count: Usage tracking (read-only)

Using Agents in Chat

Selecting an Agent

1. In chat, select the agent from a dropdown or menu
2. Agent's configuration loads automatically
3. Send messages—agent processes with its configured settings
4. All messages use agent's instructions, tools, and middleware

Agent Behavior

When using an agent:

• System prompt (instructions) prepends to every conversation
• Model overrides apply (temperature, max tokens, etc.)
• Tools are available if enabled
• Middleware processes messages
• Nodes run if configured
• Prompt stack adds additional context

Agent Management

Editing Agents

1. Find agent in agents list
2. Click edit button
3. Modify any configuration
4. Save changes
5. All conversations using this agent update automatically

Deleting Agents

1. Click delete button on agent card
2. Confirm deletion
3. Agent status changes to "archived" (soft delete)
4. Agent no longer appears in lists
5. Existing conversations preserve agent configuration snapshot

Filtering and Searching

• Filter by Type: Private, Community
• Search by Name: Real-time search
• Sort Options: Newest, Most Used, Name

Agent Types

Private Agents

• Only you can see and use
• Full control over configuration
• Can delete or modify anytime

Community Agents

• Shared templates
• Managed by community or admins
• Read-only for most users
• Cloneable for customization

Agent Configuration Best Practices

Temperature Settings

Low Temperature (0.0 - 0.3):

• Code generation and review
• Data analysis and reporting
• Technical documentation
• Consistent, focused outputs

Medium Temperature (0.4 - 0.8):

• General conversation
• Balanced creativity and consistency
• Most use cases
• Default for versatile agents

High Temperature (0.9 - 2.0):

• Creative writing
• Brainstorming sessions
• Idea generation
• Varied, unpredictable outputs

When to Use Tools

Search Tool:

• Agents that need current information
• Research assistants
• News and trend analysis
• Real-world data lookup

Context Tool:

• Agents that reference conversation history
• Long-term project assistants
• Personalized recommendations
• Continuity across sessions

MCP Integration:

• External API access needed
• Third-party service integration
• Custom tool requirements
• Advanced automation

When to Use Middleware

PII Detection:

• Customer support agents
• Data handling assistants
• Compliance-sensitive tasks
• Privacy-critical applications

Model Fallback:

• High-reliability requirements
• Cost optimization (try cheap model first)
• Redundancy for critical agents
• Graceful degradation

When to Use Nodes

Reasoning Node:

• Complex problem-solving
• Multi-step analysis
• Code debugging and architecture
• Logical deduction tasks

Review Node:

• Quality assurance needed
• Error-prone tasks
• High-stakes outputs
• Self-correction important

Pro Tips

Tip 1: Create Reusable Agents for Recurring Tasks

If you frequently do code reviews, create a "Code Reviewer" agent with:

• Low temperature (0.3) for consistency
• Specific instructions checklist
• Reasoning node for step-by-step analysis
• Saved configuration you can reuse instantly

Tip 2: Use Agent Types to Organize

• Private agents for personal workflows
• Community agents for team-shared configurations

Keep your agent list organized and discoverable.

Tip 3: Start Simple, Add Complexity

Begin with basic configuration:

1. Name and instructions
2. Model selection
3. Temperature setting

Add tools, middleware, and nodes only when needed. Over-configuration can slow responses and complicate debugging.

Tip 4: Test with Different Temperatures

Same instructions with different temperatures produce different results:

• Test at 0.0, 0.5, and 1.0
• Find the sweet spot for your use case
• Document optimal settings in agent description

Tip 5: Combine Agents with Prompt Stacks

Agents provide base configuration, prompts add context:

• Agent: "Code Reviewer" with low temperature
• Prompt stack: Specific project guidelines
• Result: Consistent reviews customized per project

Tip 6: Use Tags for Discovery

Tag agents by:

• Domain: coding, writing, analysis
• Specialty: python, frontend, data-science
• Use Case: debugging, documentation, learning

Easy filtering when you have many agents.

Tip 7: Version Your Agents

When making significant changes:

• Increment version number in metadata
• Document changes in description
• Consider cloning instead of modifying for major rewrites

Helps track agent evolution and rollback if needed.