Feat: add debate work pattern. resolves #257

[zhjuzi]

Debate Agent Framework with LLM-driven Tool Execution

Description

Implementation of a complete Debate Agent Framework including a new work pattern, enhanced execution agent, and a sample application. The framework supports LLM-driven tool decision-making and flexible debate workflows, allowing for better handling of complex debate scenarios and enhancing execution capabilities through intelligent tool selection.

System Architecture

Features

Debate Work Pattern (debate_work_pattern.py)

• Complete implementation of debate flow work pattern
• Support for collaborative work among planning, critique, and execution agents
• Integrated consensus evaluation mechanism for multi-round debates
• Flexible debate history management
• Configurable round limits and consensus thresholds

EnhancedExecutingAgentTemplate (enhanced_executing_agent_template.py)

• LLM-powered tool selection and parameter generation
• Dynamic tool calling with feedback loop
• Tool description formatting injected into prompts
• Tool call history tracking for enhanced decision making
• Iterative tool execution support

Debate Agent Template (debate_agent_template.py)

• Implementation of complete debate agent template
• Configuration-driven design aligned with AgentUniverse framework
• Standardized interface implementation
• Expert framework integration support
• Flexible debate orchestration and agent coordination

Critique Agent Template (critique_agent_template.py)

• Dedicated critique agent template implementation
• Support for generating critique_score for consensus evaluation
• Structured feedback mechanism for concerns and suggestions
• JSON output parsing for consistent evaluation

Debate Flow Process

graph TD
    User[User Input] --> DebateAgent[Debate Agent Template]
    
    subgraph "Debate Work Pattern"
        DebateAgent --> Planning[Planning Agent]
        Planning --> Critique[Critique Agent]
        Critique --> |Consensus Check| Decision{Consensus Reached?}
        Decision -->|No| Planning
        Decision -->|Yes| Execution[Enhanced Executing Agent]
    end
    
    Execution -->|Tool Selection| ToolCalling[LLM Tool Calling]
    ToolCalling -->|JSON Format| Tools[Available Tools]
    Tools --> Results[Execution Results]
    Results --> Output[Final Output]

Loading

Technical Details

• Hybrid evaluation strategy to assess consensus between planning and critique agents
• LLM-generated critique_score as primary evaluation mechanism
• Rule-based evaluation using concern count as fallback strategy
• LLM-based tool selection instead of hardcoded tool calls
• Tool call chain implementation for multi-step execution
• Configuration-driven framework design pattern

Consensus Evaluation Process

1. Primary Strategy: Use LLM-generated critique_score

   • The CritiqueAgentTemplate outputs a 'critique_score' field (0-1 range)
   • This score directly represents the critic's assessment of the plan quality
   • The score is parsed from the LLM's structured JSON output

2. Fallback Strategy: Rule-based evaluation

   • If critique_score is missing, count the number of concerns
   • No concerns = complete consensus (1.0)
   • Each concern reduces consensus by 0.1 (linear relationship)

Example Application

• Complete debate agent application example provided
• Multi-tool integration (including markdown_writer_tool)
• Demonstration of full debate workflow
• Includes necessary configuration files and test code

Test Results

Checklist

• [√] Code follows project coding style
• [√] Added necessary documentation
• [√] Added type hints and docstrings
• [√] Properly implemented error handling and user feedback
• [√] Maintained compatibility with existing APIs
• [√] Improved code organization and maintainability
• [√] Provided complete example application