A pattern-driven tour of building production-grade agentic AI: the anatomy of an agent, the GenAI maturity model, the emerging interoperability stack (function calling, MCP, A2A), and the architecture and design patterns for coordination, robustness, explainability, and human interaction — grounded in a loan-processing use case that runs from a single agent to a full multi-agent system.
These pages are study notes — a chapter-by-chapter distillation of the key concepts, each paired with a short quiz. They are a companion to the book, not a replacement for it. For the full treatment, code, and diagrams, read the original.
Original work & credit
Dr. Ali Arsanjani & Juan Pablo Bustos
Enterprise AI architects writing on production-grade agentic systems; foreword by Thomas Kurian, CEO of Google Cloud.
- Title
- Agentic Architectural Patterns for Building Multi-Agent Systems
- Publisher
- Packt Publishing Ltd.
- Published
- January 2026
- ISBN
- 978-1-80602-957-0
These notes are independent, original summaries created for study purposes. All rights to the underlying work belong to Dr. Ali Arsanjani & Juan Pablo Bustos and Packt Publishing Ltd.. Please support the author by reading the full book.
Chapters
16 of manyGenAI in the Enterprise: Landscape, Maturity, and Agent Focus
Why 'context is king,' the sense–reason–plan–act anatomy of an agent, the 7-level GenAI Maturity Model, the emerging agentic stack (function calling, MCP, A2A), and the challenges of going from PoC to production.
Agent-Ready LLMs: Selection, Deployment, and Adaptation
The LLM as cognitive core (fit over benchmarks), the dimensions of model selection, context-window and small-vs-large trade-offs, deployment/serving options, and AgentOps for production reliability.
The Spectrum of LLM Adaptation for Agents: RAG to Fine-tuning
The 6-level Agentic AI Maturity Model, the RAG spectrum (vanilla/advanced/agentic), fine-tuning PEFT vs FFT, in-context learning, and the grounding practices that keep every method reliable.
Agentic AI Architecture: Components and Interactions
The hierarchy of autonomy (LLM vs workflow vs agent), the agent anatomy as architecture, direct vs stigmergic interaction, the three-layer stack (function calling, MCP, A2A), and five architectural features.
Multi-Agent Coordination Patterns
Eleven coordination patterns mapped to maturity — Agent Router, Supervisor vs Swarm delegation, composition topologies (Blackboard/Contract-Net/Supervision Trees), planning, knowledge sharing, and the disagreement protocols.
Explainability and Compliance Agentic Patterns
Fighting instruction drift and silent failures with four patterns: Instruction Fidelity Auditing, Fractal Chain-of-Thought Embedding, Persistent Instruction Anchoring, and Shared Epistemic Memory.
Robustness and Fault Tolerance Patterns
The five-level robustness spectrum and sixteen patterns — recovery and consensus safety nets (Watchdog, Majority Voting, Fallback Model) plus security and observability (Causal Dependency Graph, Mesh Defense, sandboxing).
Human-Agent Interaction Patterns
Five interaction patterns sorted by who initiates — Human Calls/Delegates to Agent, and Agent Calls Human (HITL escalation), Agent Delegates to Agent, and Agent Calls Proxy Agent — all in service of user trust.
Agent-Level Patterns
The incremental single-agent capability stack: Single Agent Baseline, short/long-term memory, RAG grounding, structured reasoning with self-correction (CoT + self-critique), and multimodal sensory input.
System-Level Patterns for Production Readiness
The 'city' the agent inhabits — four patterns for the non-functional requirements: Tool and Agent Registry, Agent Authentication & Authorization, Real-Time Compliance Monitoring, and Event-Driven Reactivity.
Advanced Adaptation: Building Agents That Learn
The Self-Improvement Flywheel (generate→evaluate→learn→deploy), the R⁵ operational model, the Hybrid Planner+Scorer architecture, synthetic data, coevolved training and red teaming, plus tokenomics and ROI.
A Practical Roadmap: Implementing Agentic Patterns by Maturity
A three-level implementation roadmap — foundational monolith, decoupled production service, self-improving ecosystem — and the pattern-first philosophy of budgeting focus by maturity level.
Use Case: A Single Agent for Loan Processing
Building a Level-3 monolithic loan agent with Google's ADK and Gemini — FCoT-driven cognition, FunctionTool as the agent's hands, and ThinkingConfig for observable reasoning.
Use Case: A Multi-Agent System for Loan Processing
Evolving the brittle monolith into a Level-4 multi-agent system — Supervisor + Agent Delegates to Agent + FCoT + data contracts — for fault isolation, modularity, and an explainable audit trail.
Agent Frameworks — A Multi-Agent System for Loan Processing
The same system in three frameworks — Google ADK (production lifecycle), CrewAI (team of specialists), and LangGraph (state machine) — plus observability and governance via LangSmith and OpenTelemetry.
Conclusion: Charting Your Agentic AI Journey
The capstone — a pattern-first recap, how patterns compose to reach higher maturity (a worked PTO example), and a hands-on practitioner's action plan to go from theory to a real first project.