Skip to content

02 — Agentic AI Architectures

What This Section Teaches

How multiple agents coordinate to solve problems too complex for a single agent. Each pattern is implemented in all four frameworks using the same mock task — making the cross-framework comparison explicit.

The Patterns

01 SEQUENTIAL           02 PARALLEL              03 HIERARCHICAL
─────────────           ───────────              ───────────────
[A]→[B]→[C]            [O]→[A1]                 [Manager]
                             →[A2] →[Merge]       ↙        ↘
                             →[A3]              [Lead1]   [Lead2]
                                                ↙  ↘      ↙   ↘
                                              [W1][W2]  [W3] [W4]

04 ORCHESTRATOR          05 PIPELINE             06 ADVERSARIAL
───────────────          ───────────             ──────────────
    [Plan]               [Extract]              [Proposer]
   /  |  \                  ↓                       ↓
[S1][S2][S3]            [Transform]            [Critic]
                            ↓                       ↓
                         [Load]                  [Judge]

07 REFLEXION
────────────
[Agent] → [Evaluator] → score < threshold → [Agent] (retry)
                      → score ≥ threshold → done

Pattern Reference

# Pattern When to Use Key Mechanism
01 Sequential Tasks with strict ordering/dependencies Output of one = input of next
02 Parallel Independent subtasks, need speedup Fan-out + aggregation
03 Hierarchical Large tasks with nested decomposition Multi-level orchestration
04 Orchestrator-Subagent Dynamic task delegation Planner decides who does what
05 Pipeline Pure data transformation chains Stateless stages, no coordination
06 Adversarial/Debate Need critical analysis or red-teaming Opposing agents + judge
07 Reflexion Quality-sensitive output, iterative improvement Self-critique + retry loop

The Mock Task (same across all frameworks per pattern)

Each pattern uses a simple, consistent task so you can focus on the architecture, not the domain:

Pattern Mock Task
Sequential Research topic → Summarize → Format as report
Parallel 3 agents each research a different angle, 1 merges
Hierarchical PM assigns to Dev subteam + QA subteam
Orchestrator Planner delegates to Search agent, Code agent, Write agent
Pipeline Extract raw data → Transform → Load into structured output
Adversarial Proposer argues claim, Critic refutes, Judge decides
Reflexion Agent writes answer, Evaluator scores, Agent retries if < 8/10

Folder Structure

02-Architectures/
├── 01-Sequential/
│   ├── README.md          pattern explanation + ASCII diagram
│   ├── LangChain/         sequential.ipynb + sequential.py
│   ├── LangGraph/
│   ├── CrewAI/
│   └── ADK/
├── 02-Parallel/           same structure
├── 03-Hierarchical/
├── 04-Orchestrator-Subagent/
├── 05-Pipeline/
├── 06-Adversarial-Debate/
└── 07-Reflexion/
  1. Start with Sequential (simplest coordination pattern)
  2. Move to Parallel (introduces concurrency)
  3. Then Orchestrator-Subagent (the most common production pattern)
  4. Work through the rest in order
  5. For each pattern, read all four framework implementations side-by-side