Spaces:
Sleeping
Sleeping
A newer version of the Gradio SDK is available:
5.44.1
Day 2 Part 5: Workflow Design Patterns Summary
5 Anthropic Workflow Design Patterns
1. Prompt Chaining
- Pattern: Sequential LLM calls with optional code between steps
- Architecture:
LLM β [Code] β LLM β [Code] β LLM
- Use Case: Decompose complex tasks into fixed subtasks
- Example: Sector β Pain Point β Solution
- Key: Each LLM call precisely framed for optimal response
2. Routing
- Pattern: LLM router decides which specialist model handles task
- Architecture:
Input β Router LLM β Specialist LLM (1/2/3)
- Use Case: Separation of concerns with expert models
- Key: Router classifies tasks and routes to appropriate specialists
3. Parallelization
- Pattern: Code breaks task into parallel pieces, sends to multiple LLMs
- Architecture:
Code β [LLM1, LLM2, LLM3] β Code (aggregator)
- Use Case: Concurrent subtasks or multiple attempts at same task
- Key: Code orchestrates, not LLM; can aggregate results
4. Orchestrator-Worker
- Pattern: LLM breaks down complex task, other LLMs execute, LLM recombines
- Architecture:
Orchestrator LLM β [Worker LLMs] β Orchestrator LLM
- Use Case: Dynamic task decomposition and synthesis
- Key: LLM (not code) does orchestration; more flexible than parallelization
5. Evaluator-Optimizer
- Pattern: Generator LLM creates solution, Evaluator LLM validates/rejects
- Architecture:
Generator LLM β Evaluator LLM β [Accept/Reject Loop]
- Use Case: Quality assurance and accuracy improvement
- Key: Feedback loop for validation; most commonly used pattern
Key Architectural Insights
- Blurred Lines: Distinction between workflows and agents is artificial
- Autonomy Elements: Even workflows can have discretion and autonomy
- Guardrails: Workflows provide constraints while maintaining flexibility
- Production Focus: Evaluator pattern crucial for accuracy and robustness
Pattern Comparison
Pattern | Orchestrator | Flexibility | Use Case |
---|---|---|---|
Prompt Chaining | Code | Low | Sequential tasks |
Routing | LLM | Medium | Expert selection |
Parallelization | Code | Medium | Concurrent tasks |
Orchestrator-Worker | LLM | High | Dynamic decomposition |
Evaluator-Optimizer | LLM | High | Quality assurance |