Loading...
👨💼
Supervisor-Worker Pattern(SVW)
Orchestrator-worker architecture where a lead agent coordinates specialized subagents for parallel task execution
Complexity: highMulti-Agent
🎯 30-Second Overview
Pattern: Orchestrator-worker architecture where a lead agent coordinates specialized subagents for parallel task execution
Why: Achieves 90% performance improvement through parallel exploration and specialized expertise while maintaining centralized quality control
Key Insight: Dynamic task decomposition + separate worker contexts + real-time coordination = superior research quality at 15x token cost
⚡ Quick Implementation
1Decompose:Break query into 3-5 parallel subtasks
2Spawn:Create specialized worker agents dynamically
3Coordinate:Monitor progress across separate contexts
4Adapt:Spawn additional workers if needed
5Synthesize:Aggregate all findings into final result
Example: query → decompose(4_tasks) → spawn_workers() → monitor_parallel() → synthesize_results()
📋 Do's & Don'ts
✅Use separate context windows for each worker agent
✅Implement dynamic worker spawning based on complexity
✅Monitor worker progress and adjust strategies in real-time
✅Set clear objectives and success criteria for each worker
✅Use interleaved thinking for worker self-evaluation
✅Implement timeout and failure handling for workers
✅Cache expensive operations and intermediate results
✅Use structured output formats for worker communication
❌Create too many workers (diminishing returns after 5-7)
❌Share context between workers (reduces parallel benefits)
❌Ignore token consumption explosion (15x normal usage)
❌Skip quality control in final synthesis step
❌Use for simple queries that don't need decomposition
❌Forget to implement worker error propagation handling
🚦 When to Use
Use When
- • Complex, multi-domain research queries
- • Open-ended analysis requiring multiple perspectives
- • Tasks benefiting from parallel exploration
- • Research-intensive workflows
- • High-accuracy requirements worth extra cost
- • Problems requiring diverse expertise areas
Avoid When
- • Simple, single-domain questions
- • Cost-sensitive applications (15x token usage)
- • Real-time/low-latency requirements
- • Well-defined procedural tasks
- • Limited API quota scenarios
- • Sequential dependency workflows
📊 Key Metrics
Research Quality
90% improvement over single-agent
Token Consumption
15x baseline usage (cost planning)
Parallel Efficiency
% workers completing successfully
Task Decomposition
Optimal subtask count (3-5 workers)
Synthesis Quality
Coherence of final aggregated result
Worker Utilization
% workers contributing unique value
Error Recovery
% failed workers handled gracefully
💡 Top Use Cases
Research Analysis: "Impact of AI on healthcare" → Medical, Regulatory, Economic, Ethics workers → Comprehensive report
Market Intelligence: "Competitor analysis" → Product, Financial, Strategy, Technology workers → Strategic insights
Scientific Literature: "Climate change solutions" → Physics, Policy, Engineering, Economics workers → Multi-disciplinary review
Investment Due Diligence: "Company evaluation" → Financial, Market, Risk, Technical workers → Investment recommendation
Policy Research: "Education reform" → Academic, Economic, Social, Implementation workers → Policy framework
Technology Assessment: "Blockchain adoption" → Technical, Business, Legal, Social workers → Adoption strategy
Crisis Analysis: "Supply chain disruption" → Logistics, Economic, Geopolitical, Risk workers → Response plan
Product Strategy: "Market entry analysis" → Customer, Competitive, Regulatory, Technical workers → Go-to-market plan
References & Further Reading
Deepen your understanding with these curated resources
Contribute to this collection
Know a great resource? Submit a pull request to add it.
👨💼
Supervisor-Worker Pattern(SVW)
Orchestrator-worker architecture where a lead agent coordinates specialized subagents for parallel task execution
Complexity: highMulti-Agent
🎯 30-Second Overview
Pattern: Orchestrator-worker architecture where a lead agent coordinates specialized subagents for parallel task execution
Why: Achieves 90% performance improvement through parallel exploration and specialized expertise while maintaining centralized quality control
Key Insight: Dynamic task decomposition + separate worker contexts + real-time coordination = superior research quality at 15x token cost
⚡ Quick Implementation
1Decompose:Break query into 3-5 parallel subtasks
2Spawn:Create specialized worker agents dynamically
3Coordinate:Monitor progress across separate contexts
4Adapt:Spawn additional workers if needed
5Synthesize:Aggregate all findings into final result
Example: query → decompose(4_tasks) → spawn_workers() → monitor_parallel() → synthesize_results()
📋 Do's & Don'ts
✅Use separate context windows for each worker agent
✅Implement dynamic worker spawning based on complexity
✅Monitor worker progress and adjust strategies in real-time
✅Set clear objectives and success criteria for each worker
✅Use interleaved thinking for worker self-evaluation
✅Implement timeout and failure handling for workers
✅Cache expensive operations and intermediate results
✅Use structured output formats for worker communication
❌Create too many workers (diminishing returns after 5-7)
❌Share context between workers (reduces parallel benefits)
❌Ignore token consumption explosion (15x normal usage)
❌Skip quality control in final synthesis step
❌Use for simple queries that don't need decomposition
❌Forget to implement worker error propagation handling
🚦 When to Use
Use When
- • Complex, multi-domain research queries
- • Open-ended analysis requiring multiple perspectives
- • Tasks benefiting from parallel exploration
- • Research-intensive workflows
- • High-accuracy requirements worth extra cost
- • Problems requiring diverse expertise areas
Avoid When
- • Simple, single-domain questions
- • Cost-sensitive applications (15x token usage)
- • Real-time/low-latency requirements
- • Well-defined procedural tasks
- • Limited API quota scenarios
- • Sequential dependency workflows
📊 Key Metrics
Research Quality
90% improvement over single-agent
Token Consumption
15x baseline usage (cost planning)
Parallel Efficiency
% workers completing successfully
Task Decomposition
Optimal subtask count (3-5 workers)
Synthesis Quality
Coherence of final aggregated result
Worker Utilization
% workers contributing unique value
Error Recovery
% failed workers handled gracefully
💡 Top Use Cases
Research Analysis: "Impact of AI on healthcare" → Medical, Regulatory, Economic, Ethics workers → Comprehensive report
Market Intelligence: "Competitor analysis" → Product, Financial, Strategy, Technology workers → Strategic insights
Scientific Literature: "Climate change solutions" → Physics, Policy, Engineering, Economics workers → Multi-disciplinary review
Investment Due Diligence: "Company evaluation" → Financial, Market, Risk, Technical workers → Investment recommendation
Policy Research: "Education reform" → Academic, Economic, Social, Implementation workers → Policy framework
Technology Assessment: "Blockchain adoption" → Technical, Business, Legal, Social workers → Adoption strategy
Crisis Analysis: "Supply chain disruption" → Logistics, Economic, Geopolitical, Risk workers → Response plan
Product Strategy: "Market entry analysis" → Customer, Competitive, Regulatory, Technical workers → Go-to-market plan
References & Further Reading
Deepen your understanding with these curated resources
Contribute to this collection
Know a great resource? Submit a pull request to add it.