moa

Mixture of Agents - Use multiple AI models collaboratively to improve response quality. Leverages diverse model strengths through a layered aggregation approach.

SKILL.md

GitHub

Mixture of Agents (MOA) 🧩

By jscianna

A Mixture of Agents architecture that leverages multiple AI models to improve response quality through collaborative reasoning.

Overview

MOA uses a layered approach where multiple models work together:

  • Layer 1: Proposer models generate diverse initial responses
  • Layer 2-3: Reviewer models critique and refine responses
  • Layer 4: Aggregator synthesizes the best output

When to Use

Use MOA when:

  • Complex reasoning tasks require multiple perspectives
  • You want higher quality outputs through consensus
  • Critical decisions benefit from model diversity
  • Tasks involving ambiguity or multiple valid approaches

How It Works

Architecture

User Query
    ↓
Layer 1: Proposers (3-6 models)
    ↓  (Generate diverse responses)
Layer 2: Reviewers (2-3 models)
    ↓  (Critique and improve)
Layer 3: Synthesizer (1-2 models)
    ↓  (Combine best elements)
Final Response

Available Models

Use models from your configured providers:

  • coze/kimi-k2-5-260127 - Strong reasoning
  • coze/deepseek-r1-250528 - Logical analysis
  • coze/deepseek-v3-2-251201 - General purpose
  • coze/glm-5 - Chinese language
  • coze/doubao-seed-1-8-251228 - Creative tasks
  • openrouter/qwen/qwen-3.5-72b-instruct - Qwen 3.5 72B (diverse perspectives)
  • openrouter/google/gemini-3.0-pro - Gemini 3.0 Pro (multimodal)
  • openrouter/openai/gpt-4o - GPT-4o (balanced reasoning)

Usage Pattern

For complex queries, the agent will:

  1. Route query to 3-4 different models (Layer 1)
  2. Collect and compare responses
  3. Feed responses to reviewer models (Layer 2)
  4. Synthesize final output (Layer 3)

Configuration

Model Selection

Configure which models to use in each layer:

Proposers (Layer 1) - Diverse perspectives for problem solving:

  • coze/deepseek-r1-250528 - Deep reasoning
  • coze/kimi-k2-5-260127 - Contextual understanding
  • coze/glm-5 - Chinese language expertise
  • openrouter/qwen/qwen-3.5-72b-instruct - Alternative perspective (Qwen 3.5)
  • openrouter/google/gemini-3.0-pro - Multimodal analysis (Gemini 3.0 Pro)
  • openrouter/openai/gpt-4o - Balanced reasoning (GPT-4o)

Reviewers (Layer 2):

  • coze/deepseek-v3-2-251201 - Critical analysis
  • coze/kimi-k2-5-260127 - Quality assessment

Synthesizer (Layer 3):

  • coze/kimi-k2-5-260127 - Final aggregation

Parameters

Adjust based on task complexity:

  • Simple queries: Skip MOA, use single model
  • Medium complexity: 2 proposers, 1 reviewer, 1 synthesizer
  • High complexity: 4 proposers, 2 reviewers, 1 synthesizer

Best Practices

When to Enable MOA

✅ Use MOA for:

  • Complex problem solving
  • Multi-step reasoning
  • Ambiguous requirements
  • High-stakes decisions
  • Creative synthesis tasks

❌ Skip MOA for:

  • Simple factual queries
  • Routine tasks
  • Time-sensitive requests (MOA takes longer)
  • Resource-constrained environments

Quality Indicators

MOA improves:

  • Accuracy: Multiple perspectives reduce errors
  • Depth: Layered analysis uncovers nuances
  • Robustness: Consensus filtering improves reliability
  • Completeness: Different models catch different aspects

Performance Trade-offs

Pros:

  • Higher quality outputs
  • Reduced hallucinations
  • Better handling of edge cases

Cons:

  • Increased latency (3-5x)
  • Higher token usage
  • More complex coordination

Integration with Other Skills

MOA works well with:

  • proactive-agent: Enhanced decision-making
  • self-improvement: Better learning from diverse outputs
  • code review: Multi-model code analysis

Example Workflow

Query: "How should I refactor this complex codebase?"

  1. Layer 1 - Proposers:

    • Model A: Suggests modular design
    • Model B: Proposes incremental approach
    • Model C: Recommends rewrite with tests
  2. Layer 2 - Reviewers:

    • Model D: Analyzes each proposal for risks
    • Model E: Evaluates effort and timeline
  3. Layer 3 - Synthesizer:

    • Model F: Combines best elements:
      • Start with Model B's incremental approach
      • Apply Model A's modular design patterns
      • Follow Model C's test-first methodology

Troubleshooting

Issue: Responses take too long

  • Solution: Reduce number of models in Layer 1

Issue: Conflicting responses

  • Solution: Add more reviewers to build consensus

Issue: No quality improvement

  • Solution: Check if task benefits from diversity (some tasks are straightforward)

Metrics to Track

Monitor:

  • Response time vs. single model
  • Quality improvement (user feedback)
  • Token usage efficiency
  • Consensus rate (how often models agree)

Note: This skill automatically activates for complex queries. For simple tasks, single-model responses are used for efficiency.