Ensemble Problem Solving

When to Use

• "Give me options for..."
• "What's the best way to..."
• "Explore different approaches..."
• "I want to see alternatives..."
• "Compare approaches for..."
• "Which approach should I use..."

• Architecture decisions with trade-offs
• Code generation with multiple valid implementations
• API design with different philosophies
• Naming, branding, documentation style
• Refactoring strategies
• Algorithm selection

• Simple lookups or syntax questions
• Single-cause bug fixes
• File operations, git commands
• Deterministic configuration changes
• Tasks with one obvious solution

What It Does

1. Analyzes the task to determine if ensemble approach is valuable
2. Generates 3 distinct prompts using appropriate diversification strategy
3. Spawns 3 parallel subagents to develop solutions independently
4. Evaluates all solutions using weighted criteria
5. Returns the best solution with explanation and alternatives summary

Approach

Step 1: Classify Task Type

• Code Generation: Functions, classes, APIs, algorithms
• Architecture/Design: System design, data models, patterns
• Creative: Writing, naming, documentation

Step 2: Invoke Ensemble Orchestrator

Task tool with:
- subagent_type: 'ensemble-orchestrator'
- description: 'Generate and evaluate 3 parallel solutions'
- prompt: [User's original task with full context]

• Prompt diversification
• Parallel execution
• Solution evaluation
• Winner selection

Step 3: Present Result

• The winning solution (in full)
• Evaluation scores for all 3 approaches
• Why the winner was selected
• When alternatives might be preferred

Diversification Strategies

Evaluation Rubric

Example

1. Classifies as Code Generation
2. Invokes ensemble-orchestrator
3. Three approaches generated:
   • Simple: Token bucket with in-memory counter
   • Performance: Sliding window with atomic operations
   • Extensible: Strategy pattern with pluggable backends
4. Evaluation selects extensible approach (score 8.4)
5. Returns full implementation with explanation

## Selected Solution

[Full rate limiter implementation with strategy pattern]

## Why This Solution Won

The extensible approach scored highest (8.4) because it provides
a clean abstraction that works for both simple use cases and
complex distributed scenarios. The strategy pattern allows
swapping Redis/Memcached backends without code changes.

## Alternatives

- **Simple approach**: Best if you just need basic in-memory
  limiting and will never scale beyond one process.

- **Performance approach**: Best for high-throughput scenarios
  where every microsecond matters.

Success Criteria

• 3 genuinely different solutions generated
• Clear evaluation rationale provided
• Winner selected with confidence
• Alternatives summarized with use cases
• User understands trade-offs

Token Cost

• High-stakes architecture decisions
• Creative work where first attempt rarely optimal
• Learning scenarios where seeing alternatives is valuable
• Code that will be maintained long-term

Integration

• feature-planning: Can ensemble architecture decisions
• code-auditor: Can ensemble analysis perspectives
• plan-implementer: Executes the winning approach