Skip to main content

Agent System Overview

TradingAgents employs a sophisticated multi-agent architecture that mirrors the organizational structure of professional trading firms. Each agent has specialized responsibilities and contributes unique expertise to the trading decision-making process.

Agent Hierarchy​

The framework consists of four main agent categories working in a structured workflow:

Market Data Input
       ↓
┌─────────────────┐
│  Analyst Team   │ → Individual Analysis
├─────────────────┤
│ • Fundamentals  │
│ • Sentiment     │
│ • News          │
│ • Technical     │
└─────────────────┘
       ↓
┌─────────────────┐
│ Researcher Team │ → Collaborative Debate
├─────────────────┤
│ • Bullish       │
│ • Bearish       │
└─────────────────┘
       ↓
┌─────────────────┐
│  Trader Agent   │ → Strategy Formation
└─────────────────┘
       ↓
┌─────────────────┐
│ Risk Management │ → Final Decision
│ & Portfolio Mgr │
└─────────────────┘
       ↓
Trading Decision Output

Agent Roles​

Analyst Team​

The analyst team forms the foundation of market analysis:

Researcher Team​

The researcher team provides critical evaluation through structured debate:

Execution Team​

The execution team makes final decisions:

Agent Communication​

Information Flow​

Agents communicate through structured data formats:

  • Analysis reports from individual agents
  • Debate transcripts from researcher discussions
  • Risk assessments from risk management
  • Final trading recommendations

Collaboration Mechanisms​

  • Sequential Processing: Each team builds upon previous analysis
  • Parallel Analysis: Multiple analysts work simultaneously
  • Debate Resolution: Researchers engage in structured arguments
  • Consensus Building: Final decisions require multiple agent agreement

Agent Configuration​

LLM Assignment​

Different agents can use different LLM models based on their needs:

  • Deep Analysis: Complex tasks use o1-preview or gpt-4o
  • Quick Responses: Simple tasks use gpt-4o-mini
  • Cost Optimization: Testing environments use smaller models

Customization Options​

Each agent type supports customization:

  • Prompt Engineering: Customize agent personalities and focus areas
  • Data Sources: Configure which data feeds each agent accesses
  • Decision Weights: Adjust the influence of different agent types
  • Debate Parameters: Control discussion length and depth

Agent Performance​

Monitoring Capabilities​

The framework provides visibility into:

  • Individual agent analysis quality
  • Inter-agent agreement levels
  • Decision confidence scores
  • Processing time metrics

Optimization Strategies​

  • Model Selection: Choose optimal LLMs for each agent type
  • Prompt Tuning: Refine agent instructions for better performance
  • Data Quality: Ensure high-quality inputs for better analysis
  • Feedback Loops: Incorporate performance data into agent improvement

Next Steps​

Explore detailed documentation for each agent type: