Tutorial: Creating Custom Council Configurations for Specific Tasks

Why Custom Configurations Matter

Different tasks require different council compositions. A coding council needs different models than a legal analysis council.

Configuration Components

Model Selection

Choose models based on task requirements:

• Reasoning depth: Claude, GPT-4o
• Domain expertise: Specialized models
• Speed requirements: Smaller models
• Cost constraints: Mix of sizes

Consensus Mode

• Majority Vote: Fast, simple
• Peer Review: Higher accuracy
• Debate: Best for complex issues
• Mixture of Agents: Creative tasks

Parameters

• Max tokens
• Temperature
• Consensus threshold
• Timeout limits

Example Configurations

Coding Council

{
  "name": "Coding Council",
  "models": [
    "claude-3.5-sonnet",
    "gpt-4o",
    "glm-5",
    "deepseek-coder"
  ],
  "mode": "peer_review",
  "threshold": 0.8,
  "max_tokens": 2048
}

Why this works:

• Claude: Excellent code generation
• GPT-4o: Strong debugging
• GLM-5: SOTA on SWE-bench
• DeepSeek Coder: Specialized for code

Research Council

{
  "name": "Research Council",
  "models": [
    "claude-3.5-sonnet",
    "gpt-4o",
    "gemini-1.5-pro",
    "perplexity-sonar"
  ],
  "mode": "debate",
  "threshold": 0.75,
  "max_tokens": 4096
}

Why this works:

• Long context for papers (Gemini)
• Reasoning depth (Claude)
• Knowledge breadth (GPT-4o)
• Real-time search (Perplexity)

Legal Council

{
  "name": "Legal Council",
  "models": [
    "claude-3.5-sonnet",
    "gpt-4o",
    "gemini-1.5-pro"
  ],
  "mode": "debate",
  "threshold": 0.85,
  "max_tokens": 4096,
  "peer_review": true
}

High threshold because legal accuracy is critical.

Creative Council

{
  "name": "Creative Council",
  "models": [
    "claude-3.5-sonnet",
    "gpt-4o",
    "gemini-1.5-flash"
  ],
  "mode": "mixture_of_agents",
  "temperature": 0.8,
  "max_tokens": 2048
}

Higher temperature and MoA for creative diversity.

Creating Your Configuration

1. Identify your primary use case
2. Select models with complementary strengths
3. Choose appropriate consensus mode
4. Set conservative thresholds initially
5. Test and iterate

Testing Configurations

Run test queries and evaluate:

• Accuracy on known answers
• Response quality
• Latency
• Cost per query

Adjust based on results.