Agentic Architectural Patterns for Building Multi-Agent Systems

Author: Dr. Ali Arsanjani, Juan Pablo Bustos | Chapters: 16 | Generated: 2025-07-11

How to Use This Skill

• Without arguments — load core patterns for reference
• With a topic — ask about supervisor pattern, robustness, FCoT, or another indexed topic; I find and read the relevant chapter
• With chapter — ask for ch05; I load that specific chapter
• Browse — ask "what chapters do you have?" to see the full index

When you ask about a topic not covered in Core Patterns below, I will read

the relevant chapter file before answering.

Core Architectural Patterns

GenAI Maturity Model (Ch 1, 3)

Six-level progression: Basic LLM apps → RAG-enhanced → Single-agent → Multi-agent → Autonomous collectives → Self-improving. Use to assess organizational readiness and plan adoption. Don't skip levels.

Agent Anatomy: Sense-Reason-Plan-Act (Ch 1, 4)

Core loop every agent follows. Sense = perceive environment; Reason = interpret and decide; Plan = decompose goals; Act = execute via tools. An agent is NOT just an LLM call — it maintains state, pursues goals, and acts autonomously.

Supervisor/Orchestrator Pattern (Ch 5, 14)

Central manager agent plans workflow, delegates to specialist agents, monitors progress, synthesizes results. Best for regulated, auditable workflows. Specialists do one thing exceptionally well. Use when you need clear chain of command.

Swarm Architecture (Ch 5)

Decentralized peer-to-peer coordination without central orchestrator. Best for creative, exploratory tasks. More adaptive but harder to audit. Use when predictability matters less than emergence.

Fractal Chain-of-Thought (FCoT) (Ch 6, 13)

Agent outputs structured JSON "thought" objects citing data and rules at each reasoning step. Enables auditability, self-correction, and compliance verification. Use for any agent making decisions that need to be explained.

Instruction Fidelity Auditing (Ch 6)

Separate auditing agent inspects output against original instructions before finalization. Catches "silent failures" where sub-tasks succeed but overall intent is misaligned. Essential in hierarchical multi-agent systems.

Persistent Instruction Anchoring (Ch 6)

Core objectives and constraints repeated throughout task execution. Prevents instruction drift — goals being "lost in the middle" of long contexts. Use for any multi-step task.

Watchdog Timeout Supervisor (Ch 7)

Wrap all agent calls with timeout. Prevents a single hanging agent from freezing the entire system. Non-negotiable safety net for any production agent calling external APIs.

Adaptive Retry with Prompt Mutation (Ch 7)

Retry failed tasks with modified prompts — same input + same context = same error. Mutate the prompt on retry (rephrase, add context, simplify). Use for transient failures.

Agent Calls Human (Ch 8)

Human-in-the-Loop escalation. Agent pauses → notifies human with context → waits for decision → continues. Escalation is a core feature, not a failure. Use for ambiguity, high-stakes decisions, and safety boundaries.

Self-Improvement Flywheel (Ch 11)

Generate → Evaluate → Learn → Deploy. Closed-loop system for agents that improve over time. Balance exploitation (refine known strategies) with exploration (try novel approaches). Requires Hybrid Planner+Scorer Architecture.

Pattern-First Architecture (Ch 16)

Start with architectural patterns, then write prompts — not the reverse. Identify risk → Select pattern → Implement → Measure. Use this loop for every design decision.

Chapter Index

Topic Index

• Adaptive Retry → ch07
• Agent Anatomy → ch01, ch04, ch09
• Agent Authentication → ch10
• Agent Router → ch05
• AgentOps → ch02
• Canary Testing → ch07
• Coevolution → ch11
• Compliance Monitoring → ch10
• Consensus → ch05, ch07
• Context Management → ch01
• Event-Driven Architecture → ch10
• Fine-tuning / PEFT / DPO → ch03, ch11
• FCoT (Fractal Chain-of-Thought) → ch06, ch13
• Function Calling → ch01, ch04
• Grounding / Hallucination → ch01, ch03, ch09
• Human-in-the-Loop → ch08
• Instruction Drift / Anchoring → ch06
• Knowledge Sharing / Blackboard → ch05
• Maturity Model → ch01, ch03, ch12
• Multi-Agent Planning → ch05
• Negotiation → ch05
• RAG → ch03, ch09
• Robustness Levels → ch07
• Self-Improvement Flywheel → ch11
• Supervisor Pattern → ch05, ch14
• Swarm Architecture → ch05
• Tool Routing → ch05
• Watchdog Timeout → ch07

Supporting Files

• glossary.md — all key terms with definitions
• patterns.md — all architectural patterns with when-to-use and trade-offs
• cheatsheet.md — quick reference tables and decision guides

Scope & Limits

This skill covers the book content only. For hands-on implementation, combine with

framework-specific tools (Google ADK, CrewAI, LangGraph). For topics beyond this book,

check related skills or ask the agent directly.