Loading...
👥
Producer-Critic Pattern(PC)
Architectural pattern where separate producer and critic agents collaborate for quality improvement
Complexity: mediumReflection
🎯 30-Second Overview
Pattern: Separate producer and critic agents for unbiased quality improvement
Why: Prevents self-evaluation bias, enables specialized critique, improves output quality
Key Insight: Producer focuses on creation, Critic provides objective evaluation → Higher quality
⚡ Quick Implementation
1Define Roles:Producer generates, Critic evaluates
2Separate Prompts:Distinct system prompts for each agent
3Producer Executes:Initial content generation
4Critic Analyzes:Structured evaluation & feedback
5Iterate:Producer refines based on critique
Example: Producer(draft_code) → Critic(review) → Producer(refined_code)
📋 Do's & Don'ts
✅Use distinct personas (e.g., "Senior Engineer" for critic)
✅Provide structured evaluation criteria to the critic
✅Separate concerns completely between producer and critic
✅Consider different critic implementations (LLM, rules, human)
✅Implement stopping conditions to prevent infinite loops
❌Use the same prompt/persona for both roles
❌Let producer see critic instructions (bias prevention)
❌Skip validation of critic feedback quality
❌Allow unlimited iterations without convergence checks
❌Ignore computational costs of dual-agent approach
🚦 When to Use
Use When
- • High-quality content generation
- • Code review and optimization
- • Complex reasoning validation
- • Creative work refinement
Avoid When
- • Simple, straightforward tasks
- • Real-time applications
- • Limited computational budget
- • Well-defined single-pass solutions
📊 Key Metrics
Quality Delta
Improvement from v1 to final
Critique Precision
Valid issues / Total critiques
Convergence Rate
Iterations to acceptable quality
Cost Efficiency
Quality gain / Token cost
Bias Reduction
Objective evaluation score
Error Detection
% of actual errors caught
💡 Top Use Cases
Code Generation: Developer writes → Senior Engineer reviews → Developer refines
Content Creation: Writer drafts → Editor critiques → Writer polishes
Research Analysis: Analyst presents → Peer reviews → Analyst strengthens
Design Documents: Architect proposes → Tech Lead evaluates → Architect improves
LLM-as-Judge Implementation: Multiple outputs → LLM critic ranks/scores → Select best
References & Further Reading
Deepen your understanding with these curated resources
Academic Papers
Contribute to this collection
Know a great resource? Submit a pull request to add it.
👥
Producer-Critic Pattern(PC)
Architectural pattern where separate producer and critic agents collaborate for quality improvement
Complexity: mediumReflection
🎯 30-Second Overview
Pattern: Separate producer and critic agents for unbiased quality improvement
Why: Prevents self-evaluation bias, enables specialized critique, improves output quality
Key Insight: Producer focuses on creation, Critic provides objective evaluation → Higher quality
⚡ Quick Implementation
1Define Roles:Producer generates, Critic evaluates
2Separate Prompts:Distinct system prompts for each agent
3Producer Executes:Initial content generation
4Critic Analyzes:Structured evaluation & feedback
5Iterate:Producer refines based on critique
Example: Producer(draft_code) → Critic(review) → Producer(refined_code)
📋 Do's & Don'ts
✅Use distinct personas (e.g., "Senior Engineer" for critic)
✅Provide structured evaluation criteria to the critic
✅Separate concerns completely between producer and critic
✅Consider different critic implementations (LLM, rules, human)
✅Implement stopping conditions to prevent infinite loops
❌Use the same prompt/persona for both roles
❌Let producer see critic instructions (bias prevention)
❌Skip validation of critic feedback quality
❌Allow unlimited iterations without convergence checks
❌Ignore computational costs of dual-agent approach
🚦 When to Use
Use When
- • High-quality content generation
- • Code review and optimization
- • Complex reasoning validation
- • Creative work refinement
Avoid When
- • Simple, straightforward tasks
- • Real-time applications
- • Limited computational budget
- • Well-defined single-pass solutions
📊 Key Metrics
Quality Delta
Improvement from v1 to final
Critique Precision
Valid issues / Total critiques
Convergence Rate
Iterations to acceptable quality
Cost Efficiency
Quality gain / Token cost
Bias Reduction
Objective evaluation score
Error Detection
% of actual errors caught
💡 Top Use Cases
Code Generation: Developer writes → Senior Engineer reviews → Developer refines
Content Creation: Writer drafts → Editor critiques → Writer polishes
Research Analysis: Analyst presents → Peer reviews → Analyst strengthens
Design Documents: Architect proposes → Tech Lead evaluates → Architect improves
LLM-as-Judge Implementation: Multiple outputs → LLM critic ranks/scores → Select best
References & Further Reading
Deepen your understanding with these curated resources
Academic Papers
Contribute to this collection
Know a great resource? Submit a pull request to add it.