Book TOC · Chapter 0 of 8 · Next: Chapter 1

Chapter 0: The PAA Reference Architecture Stack & Core Architecture Principles

How Each Consolidated Chapter Teaches the Exam Takeaways

Every chapter in this compact edition is structured around five high-retention learning layers:

1. Domain Overview and Reference Architecture: A system-level Mermaid diagram and mental map establishing how the domain's services interact in production.
2. Foundational Primitives & Default Behaviors: A high-speed synthesis of core service primitives, default behaviors, and foundational visual mental models.
3. Consideration-by-Consideration Consolidated Takeaways: Every official Exam Guide consideration (1.1a, 1.1b, ...) taught through an Architectural Briefing, Curated Best-Fit Visual Illustrations, Takeaway Pillars (Scenario Trigger → Winning Architecture → Exam Traps to Reject), and Consolidated Golden Rules.
4. Multi-Constraint Production Architecture Patterns: High-complexity production scenarios combining security, scale, latency, and FinOps constraints.
5. Head-to-Head Trade-Off Clinics & Rapid Pattern-Trigger Cheat Sheets: Side-by-side comparison matrices and If the prompt says X → Select Y lookup tables.

1. The Master Map: All 28 In-Scope Tools of the PAA Exam

Page 4 of the official Google Cloud Professional Agentic Architect Exam Guide defines 28 tools and services that are explicitly in scope. Rather than memorizing them alphabetically, organize them in your mind across Six Architectural Planes:

The 28 In-Scope Tools at a Glance

2. The 10 Core Architecture Principles of the Professional Agentic Architect Exam

Internalize these 10 Core Architecture Principles. Whenever you are torn between two plausible-sounding options on the exam, these laws will almost always break the tie:

1. Principle 1 — Deterministic Controls Always Beat Prompt Instructions for Security: Never rely on a system prompt or workspace rule ("Never run destructive commands", "Do not query the payroll table") to enforce security. Adversarial prompt injection can bypass natural language instructions. Always choose Pre-Tool Execution Hooks, IAM / Agent Identity, Principal Access Boundary (PAB) policies, or Model Armor.
2. Principle 2 — Never Use Static Keys; Always Use Short-Lived Federated or OAuth Tokens: Any option proposing static Service Account JSON keys, hardcoded API keys in prompts/env vars, or shared Basic Auth is an immediate distractor. Choose Workload Identity Federation, Auth Manager (OAuth 2.0), or short-lived OIDC/IAM Bearer tokens.
3. Principle 3 — Externalize State on Serverless Compute: Cloud Run containers are ephemeral and scale horizontally; session affinity is best-effort and never guarantees state survival. Always externalize multi-turn checkpoints to Agent Runtime Managed Sessions, Cloud Firestore, or Memorystore for Redis—and pass large binary artifacts via Cloud Storage URIs rather than stuffing megabytes into the state object.
4. Principle 4 — Use MCP for Deterministic Tools/Resources (North–South) and A2A for Peer Reasoning Agents (East–West): Never wrap a stateless SQL query or REST API in an autonomous A2A agent (wastes tokens and latency). Conversely, when two independent, goal-directed agents across frameworks or projects must negotiate subtasks and hand off control, use Agent2Agent (A2A) with Agent Cards in Agent Registry.
5. Principle 5 — Separate Short-Term Working Session State from Long-Term Episodic/Semantic Memory: Keep active turn-by-turn scratchpad variables and sliding-window history in Managed Sessions (with strict TTLs), and asynchronously distill durable user preferences and cross-session facts into Agent Platform Memory Bank.
6. Principle 6 — Match Multi-Agent Topology to Task Dependency:
7. Zero data dependency between subtasks + latency constraint → Parallel (Scatter-Gather).
8. Strict stage-by-stage dependency ($A → B → C$) → Sequential Pipeline.
9. Single intent classification to 1 of $N$ specialists → Router / Dispatcher.
10. Generate → Critique/Test → Refine until passing → Cyclic Graph / Reflection Loop with a mandatory max iteration counter.
11. Principle 7 — Progressive Disclosure Prevents Context Bloat: Whether loading 50 runbooks in Antigravity (Agent Skills SKILL.md frontmatter) or exposing 40 APIs to an ADK agent (dynamic tool retrieval / MCP tools/list), never dump all schemas or documentation into the system prompt on every turn. And if a static context payload is legitimately huge (e.g., a 500-page legal corpus or 20MB schema queried repeatedly), enable Vertex AI Context Caching.
12. Principle 8 — Identity Propagation Must Preserve End-User ACLs: When an enterprise agent queries shared knowledge stores (Agent Search, Google Drive MCP servers, HR databases), never use a single privileged agent service account to fetch data for all users. Propagate the end-user's identity via OAuth 2.0 On-Behalf-Of (OBO) token exchange or native Agent Search ACL enforcement.
13. Principle 9 — Evaluate the Trajectory, Not Just the Final String: An agent can hallucinate the right final answer after calling the wrong tools or failing to ground in retrieved context. In CI/CD (ADK evalset) and production (Gen AI Evaluation Service), evaluate Tool Selection Precision/Recall, Trajectory In-Order Match, and the RAG Triad (Context Relevance, Groundedness/Faithfulness, Answer Relevance).
14. Principle 10 — Defense-in-Depth Requires Both Model Armor and Sensitive Data Protection: Use Model Armor to block direct/indirect prompt injections, jailbreaks, and toxic outputs, and pair it with Sensitive Data Protection (SDP) to inspect, classify, and redact/de-identify sensitive PII/PHI in prompts, RAG chunks, and tool responses.

Book TOC · Chapter 0 of 8 · Next: Chapter 1