Skip to content

How to switch from other orchestrators to ReAct pattern

This guide shows you how to migrate from SimpleToolOrchestrator or other orchestrators to the ReActOrchestrator for more structured reasoning and better debugging capabilities.

Prerequisites

  • Existing agent configuration with an orchestrator module
  • Basic understanding of YAML configuration
  • Familiarity with your current orchestrator's behavior

Replace SimpleToolOrchestrator with ReActOrchestrator

Update your agent configuration to use the ReAct pattern:

# agents/react_agent.yml
id: react-agent
modules:
  - module: xaibo.primitives.modules.llm.OpenAILLM
    id: llm
    config:
      model: gpt-4
      temperature: 0.7

  - module: xaibo.primitives.modules.tools.python_tool_provider.PythonToolProvider
    id: tools
    config:
      tool_packages: [tools.weather, tools.calendar]

  # Replace this line:
  # - module: xaibo.primitives.modules.orchestrator.SimpleToolOrchestrator
  # With this:
  - module: xaibo.primitives.modules.orchestrator.ReActOrchestrator
    id: orchestrator
    config:
      max_iterations: 10
      show_reasoning: true
      reasoning_temperature: 0.7

exchange:
  - module: orchestrator
    protocol: LLMProtocol
    provider: llm
  - module: orchestrator
    protocol: ToolProviderProtocol
    provider: tools
  - module: __entry__
    protocol: TextMessageHandlerProtocol
    provider: orchestrator

Configure equivalent behavior

Map your existing configuration to ReAct parameters:

From SimpleToolOrchestrator

# Old SimpleToolOrchestrator config
- module: xaibo.primitives.modules.orchestrator.SimpleToolOrchestrator
  id: orchestrator
  config:
    max_thoughts: 15
    system_prompt: |
      You are a helpful assistant with access to tools.
      Think step by step and explain your reasoning.

# New ReActOrchestrator config
- module: xaibo.primitives.modules.orchestrator.ReActOrchestrator
  id: orchestrator
  config:
    max_iterations: 15  # Same as max_thoughts
    show_reasoning: true  # Enable to see step-by-step thinking
    system_prompt: |
      You are a helpful assistant that follows the ReAct pattern.
      Think step by step, use tools when needed, and explain your reasoning.

Configuration mapping

SimpleToolOrchestrator ReActOrchestrator Notes
max_thoughts max_iterations Same iteration limit concept
system_prompt system_prompt Direct equivalent
Temperature stress reasoning_temperature More controlled temperature management

Test the migration

Run your agent with a complex query to verify the ReAct pattern:

# Start your agent
xaibo run agents/react_agent.yml

# Test with a multi-step question
"What's the weather in Paris and what should I pack for a 3-day trip?"

Expected ReAct output with show_reasoning: true:

🤔 **THINKING...**
💭 **THOUGHT:** I need to get weather information for Paris first, then provide packing recommendations based on the conditions.

⚡ **TAKING ACTION...**
🔧 **ACTION:** I'll check the weather in Paris.
🛠️ **EXECUTING TOOL:** get_weather with args: {"location": "Paris"}
✅ **TOOL SUCCESS:** get_weather returned: {"location": "Paris", "temperature": 18, "condition": "rainy"}

👁️ **OBSERVING RESULTS...**
🔍 **OBSERVATION:** The weather shows it's 18°C and rainy in Paris. Now I can provide appropriate packing suggestions.

🤔 **THINKING...**
💭 **THOUGHT:** I have the weather information. I can now provide a final answer with packing recommendations.

⚡ **TAKING ACTION...**
✅ **FINAL ANSWER:** Based on the weather in Paris (18°C and rainy), pack layers including a waterproof jacket, umbrella, comfortable walking shoes, and warm clothes for the cool temperature.

Disable visual reasoning for production

For production environments, disable reasoning indicators:

- module: xaibo.primitives.modules.orchestrator.ReActOrchestrator
  id: orchestrator
  config:
    max_iterations: 10
    show_reasoning: false  # Clean output for users
    reasoning_temperature: 0.7

Handle migration issues

Performance differences

ReActOrchestrator may use more LLM calls due to explicit reasoning phases:

# Optimize for fewer iterations
config:
  max_iterations: 5  # Reduce from default 10
  reasoning_temperature: 0.5  # Lower temperature for faster decisions

Tool execution changes

ReActOrchestrator executes tools differently than SimpleToolOrchestrator:

  • Tools are called during ACTION phase only
  • Each tool execution triggers an OBSERVATION phase
  • No automatic temperature increase on tool failures

Response format changes

Users will see structured reasoning with ReAct vs. direct responses:

# For user-facing applications, consider:
config:
  show_reasoning: false  # Hide internal reasoning
  system_prompt: |
    Provide direct, helpful answers. Use the ReAct pattern internally
    but present clean final answers to users.

Verify successful migration

Check that your agent:

  1. Maintains functionality - Same tool usage and response quality
  2. Shows structured reasoning - Clear thought-action-observation cycles
  3. Handles errors gracefully - Continues reasoning after tool failures
  4. Respects iteration limits - Stops at max_iterations with summary

The ReAct pattern provides more transparent and debuggable agent behavior while maintaining the same core functionality as other orchestrators.