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

Chapter 0: The PAA Reference Stack & 10 Core Principles

Welcome to the Google Cloud Certified Professional Agentic Architect (PAA) architecture and certification guide. This book takes you from a solid cloud foundation to practical mastery using clear, intermediate-level explanations, visual mental models, and concrete decision rules.

How This Guide Is Organized

Each chapter focuses on a logical group of agentic architecture domains. Instead of drowning you in raw documentation or repetitive question dumps, every section follows a simple three-step flow:

1. The Big Picture: A plain-English overview and a clean architecture diagram showing how the pieces fit together.
2. Topic-by-Topic Breakdown: Each official domain consideration (1.1a, 1.1b, etc.) is explained clearly with two custom-designed visual diagrams (one for the core concept and one for the architecture/trade-off), practical bullet points, and a short Key Takeaways box.
3. Quick Decision Cheat Sheet: A single lookup table at the end of each section showing If the scenario asks for X → Choose Y → Avoid mistake Z.

1. The Master Map: All 28 Tools in the PAA Stack

The Google Cloud agentic ecosystem spans 28 core tools and services. You don't need to memorize them as an isolated list. Instead, picture them as six layers of an enterprise agent stack:

Plain-English Guide to All 28 Tools

2. The 10 Core Architectural Principles

When evaluating competing architectural options, apply these 10 core principles:

1. Enforce security with code and IAM, never with prompts: Anyone can trick an LLM into ignoring a system prompt like "Do not delete tables." Always use IAM roles, Principal Access Boundary (PAB), Pre-Tool Execution Hooks, or Model Armor for real security.
2. Never use static keys—always use short-lived tokens: Avoid downloading service account JSON keys or hardcoding API keys. Choose Workload Identity Federation, Auth Manager (OAuth 2.0), or OIDC ID tokens.
3. Store state outside the container on serverless runtimes: Cloud Run instances start and stop at any time. Save conversation state in Agent Runtime Managed Sessions, Firestore, or Memorystore for Redis, and pass large files using Cloud Storage (gs://) URIs.
4. Use MCP for tools (North–South) and A2A for peer agents (East–West): If an agent needs to run a SQL query or call a REST API, use an MCP Server. If an agent needs to delegate a complex goal to another autonomous agent, use Agent2Agent (A2A).
5. Keep short-term session state separate from long-term memory: Use Managed Sessions for the current chat's scratchpad variables, and use Memory Bank to remember durable user facts across future conversations.
6. Match the multi-agent pattern to how tasks depend on each other:
7. Independent tasks that need speed → Parallel Agent
8. Step B needs Step A's output → Sequential Agent
9. Send request to 1 of N specialists → Router / Coordinator Agent
10. Draft, test, and fix until passing → Loop / Cyclic Graph Agent (always set a maximum iteration limit!)
11. Load context on demand (Progressive Disclosure): Don't paste 50 tool schemas or 100 pages of instructions into every prompt. Use Antigravity Skills (which load only when needed) and Vertex AI Context Caching for large, repeated reference documents.
12. Preserve the human user's permissions (ACL Pass-Through): When an agent searches company files or calls a workspace tool for an employee, pass the employee's identity using OAuth 2.0 On-Behalf-Of (OBO) or Agent Search ACLs—never use a single admin account for everyone.
13. Test the agent's steps (Trajectory), not just the final answer: Use ADK evalset in CI/CD and the Gen AI Evaluation Service to check that the agent called the right tools in the right order and stayed faithful to retrieved facts (Groundedness).
14. Combine Model Armor + Sensitive Data Protection: Use Model Armor to stop prompt injections and jailbreaks, and Sensitive Data Protection (SDP) to redact PII like credit cards or SSNs.

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