Do Competitively

功能

• Competitive generation with self-critique
• Meta-judge evaluation specification
• Multi-judge evaluation with verification loops
• Adaptive strategy selection (polish, synthesize, redesign)
• Evidence-based synthesis of results

使用场景

• Designing complex APIs or systems where multiple approaches need evaluation.
• Generating and refining code for high-stakes projects requiring maximum quality.
• Producing comprehensive documentation or technical specifications through comparative agent output.
• Tasks where consensus among multiple AI perspectives is crucial for optimal outcome.

非目标

• Executing trivial tasks where the overhead of multi-agent competition is not justified.
• Simple code generation without comparative evaluation or synthesis.
• Replacing human oversight entirely; the framework aims to augment decision-making.

工作流

1. Create reports directory.
2. Phase 1: Launch 3 generator agents and 1 meta-judge agent in parallel.
3. Meta-judge generates evaluation specification YAML.
4. Generators produce independent solutions.
5. Phase 2: Launch 3 judges in parallel, evaluating solutions against meta-judge spec.
6. Judges produce comparative reports and votes.
7. Phase 2.5: Orchestrator analyzes judge outputs to select strategy (SELECT_AND_POLISH, REDESIGN, FULL_SYNTHESIS).
8. Phase 3: Based on strategy, polish a winner, redesign fundamentally flawed solutions, or synthesize the best elements from multiple solutions.
9. Output final solution and reports.