Agent Editor

The Agent Editor provides a visual interface for creating and configuring AI agents. Define agent behavior through system prompts, select models, assign tools, and test functionality before deployment.

Overview

Building effective agents requires careful configuration of multiple components: the underlying LLM, behavioral instructions (system prompt), available tools, and operational parameters. The Agent Editor consolidates all configuration into a single interface with live testing capabilities.

Agent Configuration Options

Model Selection

Choose the LLM that powers your agent:

Model selection considerations:

• Task complexity — GPT-4 Opus for hard reasoning, GPT-3.5/Haiku for simple tasks
• Context requirements — Claude 3 supports 200K tokens, ideal for long documents
• Cost — Cheaper models (Haiku, Qwen) for high-volume scenarios
• Latency — Smaller models respond faster

System Prompt

The system prompt defines agent behavior, constraints, and output format:

You are a customer service agent for TechCorp.

Capabilities:
- Answer product questions using the knowledge base
- Check order status via the order lookup tool
- Escalate complex issues to human agents

Constraints:
- Never share customer PII
- Maintain professional, empathetic tone
- Provide sources for all factual claims

Output format:
- Start with a brief acknowledgment
- Provide the answer or action taken
- End with "Is there anything else I can help with?"

System prompt best practices:

• Be specific — Vague instructions like “be helpful” produce inconsistent behavior
• Define constraints — Explicit boundaries prevent unsafe outputs
• Provide examples — Few-shot examples in the system prompt improve quality
• Specify format — Structured output requirements (JSON, markdown) ensure parseable responses

Temperature and Parameters

Fine-tune generation behavior:

• Temperature (0-2) — Randomness in outputs

  • 0 = Deterministic, consistent responses
  • 0.7 = Balanced creativity
  • 1.5+ = High variability, creative outputs

• Top-P (0-1) — Nucleus sampling threshold

  • 1.0 = Consider all tokens
  • 0.9 = Restrict to top 90% probability mass

• Max tokens — Output length limit

  • Set based on expected response size
  • Higher = more complete responses but higher cost

• Presence/Frequency penalty — Reduce repetition

  • Use 0.3-0.5 to avoid repetitive phrasing

Tool Assignment

Assign tools from the tool library to give agents capabilities:

1. Click “Add Tool” in the configuration panel
2. Browse available tools by category
3. Select tools relevant to agent’s purpose
4. Configure tool-specific parameters (optional)

Example tool assignments:

Research Agent:

• Web search tool
• Wikipedia lookup tool
• PDF extractor tool

Coding Agent:

• File read/write tools
• Code execution sandbox
• Git operations tool

Customer Service Agent:

• Knowledge base search
• Order lookup tool
• CRM integration tool

Agents automatically learn how to use assigned tools through function calling or ReAct prompting.

Skills Assignment

Load skills from the filesystem-based skill library:

1. Navigate to the Skills tab
2. Browse discovered skills
3. Toggle skills to enable/disable
4. Skills auto-load when the agent starts

Skills provide complex multi-step capabilities that compose multiple tools. See Tools & Skills for details.

Creating an Agent

Testing Agents

Interactive Testing

The built-in test panel provides:

• Input field — Enter test queries
• Response viewer — See agent output with syntax highlighting
• Tool log — View which tools were called and with what parameters
• Performance metrics — Response time, token usage, cost per request

Testing workflow:

1. Enter a query representative of real usage
2. Run the agent
3. Evaluate output quality and tool usage
4. Adjust system prompt or tool assignment
5. Re-test until behavior meets requirements

Edge Case Testing

Test agent behavior in error scenarios:

• Missing information — Query that requires unavailable data
• Tool failure — Simulate tool errors to test recovery
• Ambiguous input — Unclear requests to test clarification behavior
• Security bypass attempts — Try to make agent violate constraints

Robust agents gracefully handle edge cases without breaking character or exposing vulnerabilities.

Batch Testing

For systematic evaluation:

1. Create a test suite with input/expected output pairs
2. Run the agent against all test cases
3. Calculate success rate
4. Identify failure patterns
5. Refine configuration to address failures

Batch testing enables regression detection when updating agent configuration.

Advanced Features

Agent Cloning

Duplicate existing agents to experiment with variations:

1. Select an agent from the repository
2. Click “Clone”
3. New agent inherits all configuration
4. Modify system prompt or tools
5. Compare performance against original

Cloning accelerates A/B testing of agent configurations.

Version History

All configuration changes are versioned:

• View commit history for an agent
• Diff between versions
• Rollback to previous configurations
• Tag stable versions for production

Version control ensures safe experimentation and easy rollback.

Agent Templates

Start from pre-built templates for common use cases:

• Research Assistant — Web search, PDF extraction, summarization
• Code Assistant — File operations, code execution, Git
• Data Analyst — Database query, visualization, statistical tools
• Customer Support — Knowledge base search, CRM integration, ticket creation

Templates provide working baseline configurations that can be customized.

Configuration Best Practices

System Prompt Design

Do:

• Define clear role and purpose
• Specify explicit constraints
• Provide output format requirements
• Include few-shot examples for complex tasks

Don’t:

• Use vague instructions like “be smart”
• Overload with too many responsibilities
• Assume common sense (be explicit)
• Include sensitive information in the prompt

Tool Selection

Do:

• Assign only tools the agent needs
• Order tools by usage likelihood
• Test tool combinations for conflicts
• Document tool-specific configuration

Don’t:

• Give all tools to every agent (scope creep)
• Assume the agent will figure out tool usage
• Overlook security implications of powerful tools (file write, command execution)

Temperature Tuning

Use low temperature (0-0.3) for:

• Classification tasks
• Data extraction
• Structured output generation
• Safety-critical decisions

Use high temperature (0.7-1.5) for:

• Creative writing
• Brainstorming
• Diverse option generation
• Exploration tasks

Troubleshooting

Agent Not Using Tools

Symptoms: Agent responds without calling assigned tools

Causes:

• System prompt doesn’t mention tool availability
• Temperature too low (model overly conservative)
• Tool descriptions unclear

Solutions:

• Update system prompt: “You have access to the following tools: [list]”
• Increase temperature to 0.3-0.5
• Improve tool descriptions with usage examples

Inconsistent Responses

Symptoms: Same input produces wildly different outputs

Causes:

• Temperature too high
• Ambiguous system prompt
• Conflicting tool results

Solutions:

• Lower temperature to 0-0.3
• Clarify constraints in system prompt
• Add validation logic to tool outputs

Exceeding Context Window

Symptoms: Errors about token limits

Causes:

• Long conversation history with memory enabled
• Large tool outputs
• Verbose system prompt

Solutions:

• Limit memory retention window
• Summarize tool outputs before returning to agent
• Use models with larger context (Claude 200K)

Next Steps

• Assign tools and skills to enhance agent capabilities
• Enable group memory for persistent context
• Orchestrate multiple agents for complex workflows