Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.atthene.com/llms.txt

Use this file to discover all available pages before exploring further.

The Atthene Multi-Agent System provides different agent types, each optimized for specific use cases and interaction patterns. This guide helps you choose the right agent type for your needs.

​
Available Agent Types

​
LLM Agent

Type identifier: llm_agent The most basic agent type, powered by a large language model. Best for conversational interactions and tasks that don’t require external tools or complex reasoning patterns.

Best For

  • Conversational assistants
  • Content generation
  • Question answering
  • Text summarization
  • General chat interactions

Capabilities

  • ✅ Streaming support
  • ✅ Knowledge base integration
  • ✅ Memory support
  • ❌ No tool calling
  • Configurable LLM parameters
  • Multi-turn conversations

​
Configuration Example

agents:
  - name: "assistant"
    agent_type: "llm_agent"
    description: "A helpful conversational assistant"
    
    prompt_config:
      system_prompt: |
        You are a helpful AI assistant. Provide clear, accurate, 
        and friendly responses to user questions.
    
    llm_config:
      model: "gpt-4o"
      temperature: 0.7
      max_tokens: 1000
    
    streaming_config:
      enable_streaming: true
      show_output_to_user: true

​
When to Use

Use LLM agents when you need straightforward conversational interactions without external tool access
Ideal for customer support, content creation, or general Q&A scenarios

​
ReAct Agent

Type identifier: react_agent Implements the ReAct (Reasoning + Acting) pattern, enabling agents to use tools and external resources. The agent reasons about what actions to take, executes tools, and incorporates results into its responses.

Best For

  • Web search and research
  • Data retrieval and analysis
  • Multi-step problem solving
  • Tasks requiring external information
  • Any task needing tool integration

Capabilities

  • ✅ Tool calling and execution
  • ✅ Knowledge base integration
  • ✅ Memory support
  • ❌ Streaming disabled
  • Iterative reasoning loops
  • Thought process visibility

​
Configuration Example

agents:
  - name: "research_assistant"
    agent_type: "react_agent"
    description: "AI assistant with research and analysis capabilities"
    
    prompt_config:
      system_prompt: |
        You are a research assistant with access to web search and calculation tools.
        
        When solving problems:
        1. Break down the task into steps
        2. Use tools to gather information
        3. Analyze the results
        4. Provide a comprehensive answer
    
    tools:
      - "tavily_search"
      - "tavily_extract"
      - "calculator"
    
    max_iterations: 10
    
    llm_config:
      model: "gpt-4o"
      temperature: 0.1
      max_tokens: 2000
    
    streaming_config:
      enable_streaming: true
      show_output_to_user: true
      show_reasoning: true
      show_tool_to_user: true

​
ReAct Pattern Flow

User Query → Agent Reasoning → Tool Selection → Tool Execution → 
Result Analysis → More Tools? → Final Response

​
When to Use

Use ReAct agents when tasks require external information or computations
Perfect for research, data analysis, fact-checking, or complex problem-solving
Set appropriate max_iterations to prevent infinite reasoning loops. Recommended: 5-15 iterations.

​
Supervisor Agent

Type identifier: supervisor Coordinates multiple specialized worker agents, routing tasks to the most appropriate agent and managing multi-agent workflows. Acts as an intelligent orchestrator.

Best For

  • Multi-domain systems
  • Complex workflows
  • Task delegation
  • Team coordination
  • Specialized agent teams

Capabilities

  • ✅ Streaming support
  • ✅ Tool calling (for handoffs)
  • ✅ Memory support
  • Dynamic agent routing
  • Agent handoffs
  • Workflow coordination

​
Configuration Example

agents:
  - name: "team_supervisor"
    agent_type: "supervisor"
    description: "Coordinates specialized team members"
    
    # Define which agents this supervisor manages
    supervised_agents: 
      - "research_agent"
      - "analysis_agent"
      - "writer_agent"
    
    # Control coordination flow
    return_to_supervisor: true
    max_handoffs: 15
    
    prompt_config:
      system_prompt: |
        You are a team supervisor coordinating specialized agents.
        
        Your team:
        - research_agent: Gathers information via web search
        - analysis_agent: Analyzes data and identifies patterns
        - writer_agent: Creates polished written content
        
        INSTRUCTIONS:
        - Immediately delegate tasks to the appropriate specialist
        - Use transfer tools to hand off work
        - Provide clear context when transferring
        - Evaluate results and decide next steps
    
    streaming_config:
      enable_streaming: true
      show_output_to_user: true

​
Supervisor Coordination Patterns

Iterative Coordination (return_to_supervisor: true, the default): 
User → Supervisor → Worker 1 → Supervisor → Worker 2 → Supervisor → User
Supervisor evaluates each step and orchestrates multi-agent workflows. One-Way Delegation (return_to_supervisor: false): 
User → Supervisor → Worker Agent → User
Simple routing where workers respond directly to users.

​
When to Use

Use supervisors when you have multiple specialized agents that need coordination
Ideal for complex systems requiring task decomposition and delegation
Combine different agent types under a supervisor: use ReAct agents for research, LLM agents for synthesis.

​
Agent Type Comparison

FeatureLLM AgentReAct AgentSupervisor
Tool Calling❌ No✅ Yes✅ Yes (handoffs)
Streaming✅ Yes❌ No✅ Yes
Knowledge Base✅ Yes✅ Yes✅ Yes
Memory✅ Yes✅ Yes✅ Yes
Reasoning Loop❌ Single pass✅ Iterative✅ Coordination
Multi-Agent❌ No❌ No✅ Yes
ComplexityLowMediumHigh
Best Use CaseChatResearchOrchestration

​
Choosing the Right Agent Type

​
Decision Tree

1

Do you need multiple specialized agents?

Yes → Use Supervisor AgentNo → Continue to next question
2

Do you need external tools or data?

Yes → Use ReAct AgentNo → Continue to next question
3

Do you need basic conversation?

Yes → Use LLM Agent

​
Use Case Examples

Recommended: llm_agentSimple conversational interface for answering common questions. No external tools needed.
Recommended: react_agentNeeds web search, knowledge base access, and multi-step reasoning to gather and synthesize information.
Recommended: supervisor + multiple llm_agent workersSupervisor coordinates research, writing, and editing agents for comprehensive content creation.
Recommended: react_agent with calculator and database toolsPerforms calculations, queries databases, and generates insights from data.
Recommended: supervisor + specialized react_agent workersRoutes technical, business, or creative questions to domain-specific agents with appropriate tools.

​
Configuration Best Practices

​
LLM Agent Best Practices

Temperature: Use 0.7-1.0 for creative tasks, 0.1-0.3 for factual responses
System Prompt: Be specific about tone, style, and response format

​
ReAct Agent Best Practices

Max Iterations: Set to 5-10 for most tasks. Higher values (15+) only for complex research.
Tool Selection: Only include tools the agent actually needs. More tools = more complexity.
Show Reasoning: Enable show_reasoning: true during development to debug tool usage patterns.

​
Supervisor Best Practices

Clear Delegation: Write explicit instructions in prompt_config.system_prompt about when to use each worker.
Return Strategy: Use return_to_supervisor: true (the default) for quality control and multi-step workflows. Set to false for simple one-way delegation.
Max Handoffs: Set reasonable limits (10-20) to prevent infinite coordination loops.

​
Advanced Patterns

​
Hybrid Teams

Combine multiple agent types for sophisticated systems: 
agents:
  # Supervisor coordinates the team
  - name: "supervisor"
    agent_type: "supervisor"
    supervised_agents: ["researcher", "analyst", "writer"]
  
  # ReAct agent for research
  - name: "researcher"
    agent_type: "react_agent"
    tools: ["tavily_search", "tavily_extract"]
  
  # ReAct agent for analysis
  - name: "analyst"
    agent_type: "react_agent"
    tools: ["calculator"]
  
  # LLM agent for final synthesis
  - name: "writer"
    agent_type: "llm_agent"
    llm_config:
      temperature: 0.8  # More creative for writing

​
Specialized Workers

Create domain-specific agents by customizing system prompts: 
- name: "technical_expert"
  agent_type: "react_agent"
  prompt_config:
    system_prompt: |
      You are a senior software engineer specializing in Python and system architecture.
      Provide technical, detailed responses with code examples.
  tools: ["tavily_search"]

- name: "business_expert"
  agent_type: "llm_agent"
  prompt_config:
    system_prompt: |
      You are a business strategy consultant with MBA-level expertise.
      Focus on ROI, market analysis, and strategic recommendations.

​
Next Steps

YAML Configuration

Learn the complete YAML configuration syntax

Agent Capabilities

Explore tools, streaming, and other capabilities