Tree Of Thoughts

Features

• Systematic exploration of solution space
• Pruning of unpromising branches
• Expansion of viable approaches
• Meta-judge evaluation specifications
• Multi-agent evaluation and synthesis

Use Cases

• Tackling complex coding challenges requiring exploration of different architectural patterns.
• Designing novel algorithms where multiple approaches need to be evaluated.
• Synthesizing best solutions from diverse ideas in research and development.
• Generating detailed technical specifications through iterative refinement and evaluation.

Non-Goals

• Performing simple, direct tasks that do not benefit from multi-path exploration.
• Replacing human decision-making entirely; acts as an AI assistant.
• Providing generic code generation without a structured reasoning process.

Workflow

1. Phase 1: Exploration (Propose Approaches) by 3 independent agents.
2. Phase 1.5: Dispatch Pruning Meta-Judge to generate criteria.
3. Phase 2: Pruning (Vote for Top 3 Candidates) by 3 independent judges.
4. Phase 2b: Select Top 3 Proposals based on aggregated votes.
5. Phase 3: Expansion (Develop Full Solutions) by 3 independent agents.
6. Phase 3.5: Dispatch Evaluation Meta-Judge to generate criteria.
7. Phase 4: Evaluation (Judge Full Solutions) by 3 independent judges.
8. Phase 4.5: Adaptive Strategy Selection (SELECT_AND_POLISH, REDESIGN, or FULL_SYNTHESIS).
9. Phase 5: Synthesis (if FULL_SYNTHESIS selected) - Combine best elements.

Practices

• Agent Orchestration
• Structured Reasoning
• Meta-Evaluation