Agentic Workflow Patterns

A visual guide to the five core patterns for building intelligent AI systems

Pattern 1

Prompt Chaining

Breaking complex tasks into sequential, dependent steps where each step's output feeds the next

Input Task
→
Step 1
Agent A
→
Step 2
Agent B
→
Step 3
Agent C
→
Final Output

How It Works

  • Decompose: Break complex task into smaller, manageable subtasks
  • Chain: Each step involves an LLM call with a focused prompt
  • Connect: Output of one step becomes input to the next
  • Validate: Check quality at intermediate points to prevent error propagation

Key Benefits

Enhanced Reasoning: Each step can focus deeply on a specific subtask without being overwhelmed by the full complexity.

Reduced Hallucinations: Focused attention on smaller problems leads to more accurate outputs.

Better Debuggability: Isolated steps make it easier to identify and fix issues.

⚠️ Critical Challenge: Error Propagation

Errors in early steps compound through the chain. Solution: Implement validation checkpoints between steps using programmatic checks, LLM-based evaluation, or rule-based validation.

When to Use

  • Task requires multiple sequential, dependent steps
  • Each step needs focused attention on a specific aspect
  • Complex problem can be decomposed linearly
  • Quality validation at each step is important
Real-World Example
Research Report Generation:
  1. Step 1: Extract key topics from source documents
  2. Step 2: Generate outline based on topics
  3. Step 3: Write detailed sections using outline
  4. Step 4: Create executive summary from sections
  5. Step 5: Format and polish final report
Pattern 2

Routing

Intelligently directing tasks to the most appropriate specialized agent based on classification

Incoming Task
↓
Router Agent
(Classifier)
↓
Agent A
(Type 1)
Agent B
(Type 2)
Agent C
(Type 3)

How It Works

Stage 1 - Classification: Router agent analyzes the incoming task to determine its type, category, intent, or complexity level.

Stage 2 - Dispatch: Based on classification, the task is directed to the specialized agent best suited to handle it.

Classification Methods

Rule-Based: Use keyword matching, regex patterns, or conditional logic for deterministic routing.

LLM-Based: Let an LLM understand context and categorize based on semantic meaning for more nuanced classification.

✅ Key Benefits
  • Task Specialization: Each agent optimized for specific request types
  • Resource Optimization: Route simple tasks to fast/cheap models, complex to powerful ones
  • Scalability: Easily add new specialized agents without changing core logic

When to Use

  • Handling diverse types of requests that need different handling
  • Different tasks require specialized agents or approaches
  • Resource optimization is important (cost/speed tradeoffs)
  • System needs to scale with new capabilities over time
Real-World Example
Customer Support System: Router classifies incoming messages and directs to specialized teams:
  • Billing Agent: Handles payment, invoices, refunds
  • Technical Agent: Solves product bugs, integration issues
  • Sales Agent: Answers feature questions, upgrades
  • Escalation Agent: Routes complex/angry customers to humans
Pattern 3

Parallelization

Multiple agents working simultaneously on independent subtasks, then consolidating results

Input Task
↓
Scatter
↓
Agent 1
Subtask A
Agent 2
Subtask B
Agent 3
Subtask C
↓
Gather & Synthesize
↓
Final Output

How It Works

  • Scatter: Distribute workload to multiple agents that work concurrently
  • Process: Agents execute their independent subtasks simultaneously
  • Gather: Collect all results when agents complete their work
  • Synthesize: Consolidate outputs into coherent final result

Synthesis Methods

Consolidation: Merge all outputs into a unified document (e.g., combining research from different sources).

Voting: Agents "vote" on the best answer, with majority or consensus determining output.

Debate: Agents present different perspectives, then a final agent synthesizes the discussion.

✅ Key Benefits
  • Speed: Dramatic reduction in total processing time
  • Multiple Perspectives: Different agents bring diverse approaches
  • Reliability: Consensus reduces individual agent errors

When to Use

  • Subtasks are independent and don't rely on each other
  • Speed improvement is valuable (latency reduction)
  • Multiple perspectives enhance quality or creativity
  • Consensus or voting improves reliability
Real-World Example
Market Research Report:
  • Agent 1: Analyzes competitor pricing strategies
  • Agent 2: Summarizes customer reviews and sentiment
  • Agent 3: Identifies industry trends and forecasts
  • Synthesis Agent: Combines findings into comprehensive report with cross-referenced insights
Pattern 4

Evaluator-Optimizer

Iterative refinement through evaluate-critique-improve loops until quality standards are met

Input Task
↓
Generator Agent
Creates Output
↓
Evaluator Agent
Assesses Quality
↓
Quality Check
✓ Pass
Final Output
✗ Fail
Refine & Retry ↑

How It Works

  • Generate: Generator agent creates initial output
  • Evaluate: Evaluator agent assesses quality against criteria
  • Decide: Check if output meets standards
  • Refine: If not, provide feedback and regenerate
  • Loop: Repeat until quality threshold met or max iterations reached

Evaluation Criteria Types

Task-Specific: Accuracy, completeness, relevance to requirements

Quality Attributes: Clarity, grammar, tone, formatting

Constraints: Length limits, keyword inclusion, structural requirements

⚠️ Critical: Define Stopping Conditions

Always set maximum iterations (e.g., 3-5 attempts) to prevent infinite loops. Balance quality improvement against computational cost.

When to Use

  • High-quality output is critical and worth iteration
  • Complex requirements must be satisfied
  • Iterative refinement demonstrably adds value
  • Self-correction capability improves reliability
Real-World Example
Blog Post Creation:
  1. Generator: Creates draft blog post on specified topic
  2. Evaluator: Checks for: SEO keywords, readability score, structure, tone consistency, factual accuracy
  3. Feedback Loop: "Missing keyword 'automation' in first 100 words. Intro lacks hook. Technical jargon too dense for target audience."
  4. Refinement: Generator revises based on specific critiques
  5. Iteration: Process repeats until all quality criteria satisfied
Pattern 5

Orchestrator-Workers

Central planning agent dynamically coordinates specialized worker agents to solve complex, unpredictable problems

Complex Task
↓
Orchestrator Agent
(Planning & Coordination)
↓
Dynamic Task Assignment
Worker A
Worker B
Worker C
Worker D
Results & Feedback ↑↓
Orchestrator
Adapts Plan
↓
Final Solution

How It Works

  • Analyze: Orchestrator assesses the complex problem
  • Plan: Creates strategy and breaks down into subtasks
  • Delegate: Assigns subtasks to appropriate specialist workers
  • Monitor: Tracks progress and worker outputs
  • Adapt: Adjusts plan based on results and new information
  • Iterate: Continues until problem is solved or goals achieved

Key Characteristics

Dynamic Planning: Unlike fixed workflows, the orchestrator can change its approach mid-execution based on intermediate results.

Adaptive Coordination: The orchestrator decides which workers to use, when, and in what order—adapting to the problem's evolving state.

Maximum Autonomy: This is the most "agentic" pattern, with the orchestrator making high-level strategic decisions independently.

✅ Maximum Flexibility

This pattern handles problems where you don't know the solution path in advance. The orchestrator figures it out dynamically, making it ideal for open-ended, exploratory, or highly variable tasks.

When to Use

  • Task is complex and solution path isn't known in advance
  • Problem requires dynamic planning and adaptation
  • Multiple specialized capabilities needed in unpredictable order
  • Maximum flexibility and autonomous decision-making required
Real-World Example
Business Strategy Consultant:
  • Orchestrator: Senior consultant agent analyzes "Help us enter Asian markets"
  • Initial Plan: Research → Analysis → Recommendations
  • Worker 1 (Research): Gathers market data for Asian regions
  • Orchestrator Adapts: Sees high regulatory complexity in China, assigns Worker 2 (Legal) to investigate
  • Worker 3 (Financial): Analyzes investment requirements based on legal findings
  • Orchestrator Pivots: Regulatory burden too high, shifts focus to Southeast Asia
  • Worker 4 (Competitive): Analyzes competitors in Thailand, Vietnam, Singapore
  • Final Synthesis: Orchestrator compiles strategic recommendation with adaptive insights
↑