Code Labs - 02: Architecture Patterns

Seven multi-agent coordination patterns, each implemented in all four frameworks using the same mock task. Builds directly on 01 - Agent Types.

← Back to Overview: Agentic AI · Back to Concepts: Architectural Patterns

---

Pattern Overview

Sequential          Parallel            Hierarchical
A → B → C           A ─┬─ B             Orchestrator
(pipeline)            └─ C               ├── SubAgent-1
                        ↓                └── SubAgent-2
                      Aggregate               ↓
                                        SubAgent-3

Orchestrator-       Pipeline            Adversarial         Reflexion
Subagent            (data-flow)         Debate              (self-critique)
Planner             Stage1→Stage2       Proposer            Agent
  ├── Worker-1      (no shared state)   vs Critic           ↓ critique
  └── Worker-2                          → Judge             ↓ revise
                                                            ↓ done?

---

Pattern Reference Table

#	Pattern	Key Concept	When to Use	Notes Link
01	Sequential	A→B→C pipeline, output feeds next	Linear multi-step workflows	Architectural Patterns
02	Parallel	Fan-out + aggregate	Independent subtasks, reduce latency	Architectural Patterns
03	Hierarchical	Nested orchestration tiers	Complex delegation, team-of-teams	Architectural Patterns
04	Orchestrator-Subagent	Planner + specialized workers	Dynamic task decomposition	Architectural Patterns
05	Pipeline	Data-flow, no shared state	ETL-style transformation chains	Architectural Patterns
06	Adversarial Debate	Proposer vs Critic vs Judge	High-stakes decisions, quality gates	Architectural Patterns
07	Reflexion	Self-critique + iterative revision	Output quality, agentic planning	Design Patterns

---

Files by Pattern and Framework

Each pattern lives in 06-Agentic-AI/CodeLabs/02-Architectures/<pattern-name>/<Framework>/:

Pattern	LangChain	LangGraph	CrewAI	ADK
Sequential	sequential.py + .ipynb	same	same	same
Parallel	parallel.py + .ipynb	same	same	same
Hierarchical	hierarchical.py + .ipynb	same	same	same
Orchestrator-Subagent	orchestrator.py + .ipynb	same	same	same
Pipeline	pipeline.py + .ipynb	same	same	same
Adversarial-Debate	debate.py + .ipynb	same	same	same
Reflexion	reflexion.py + .ipynb	same	same	same

28 total implementations - every pattern × every framework.

---

Recommended Order

1. Sequential - start here; simplest pattern, clear A→B→C structure
2. Parallel - fan-out to multiple agents, then aggregate results
3. Orchestrator-Subagent - dynamic planning; the most commonly used pattern in production
4. Hierarchical - nested orchestration; builds on Orchestrator-Subagent
5. Pipeline - data-flow style; useful for ETL and document processing
6. Adversarial-Debate - quality through disagreement; challenging but powerful
7. Reflexion - self-improvement loop; combines with most other patterns

---

Pattern Selection Guide

Requirement	Recommended Pattern
Fixed sequence of steps	Sequential or Pipeline
Reduce wall-clock time	Parallel
Dynamic task decomposition	Orchestrator-Subagent
Multiple teams/layers	Hierarchical
High-quality output required	Adversarial-Debate or Reflexion
Iterative refinement	Reflexion

---

Getting Started

# Start with Sequential in LangGraph (most readable)
jupyter notebook 06-Agentic-AI/CodeLabs/02-Architectures/01-Sequential/LangGraph/sequential.ipynb

# Then compare the same pattern in CrewAI
jupyter notebook 06-Agentic-AI/CodeLabs/02-Architectures/01-Sequential/CrewAI/sequential.ipynb

---

What to Read Alongside

• Architectural Patterns - conceptual deep-dive on all 8 patterns
• Design Patterns - tool-use, reflection, planning, routing
• Multi-Agent Systems - coordination, state management, failure modes

---

Next: Agentic Systems

After mastering individual patterns, see how they combine in 03 - Agentic Systems - full end-to-end production system designs.