Skip to content

Preparing for Google Professional Agentic Architect (PAA): What the Hybrid Exam Tests and How to Prepare

Nov 13, 20261 min
TL;DRGoogle's Professional Agentic Architect is assessed as a hybrid two-part exam: a proctored multiple-choice exam with Pearson, then hands-on labs on Google Skills. Both parts are required. The GA spec is a 2-hour exam, 50-60 multiple-choice questions, $200, English, one-year validity; the beta (Sept 8-30, 2026) was 3 hours, about 80 questions, $120.

Look up & save

Loading dictionary…

🌏 中文版

This is a preparation path built from official material, not an exam-day account — I have not sat this exam. Every "what it tests" points back to the official exam guide, and every "how to prepare" points to an official learning path or Google Cloud documentation. No leaked questions. Verified 2026-11-13.

Google's AI portfolio grew a new credential at the end of 2026: the Professional Agentic Architect (PAA). It is Google's newest professional-level certification, and the first official exam that treats the agentic workflow itself as the test object — it asks whether you can design, deploy, and govern autonomous agent systems, not whether you can call a model.

The exam has run on a hybrid two-part structure since the beta in September 2026 and opened for general registration on November 2, 2026 (GA). Same flow as beta: pass the Pearson-proctored multiple-choice exam, then receive hands-on labs via CM Connect on Google Skills; both parts are required to earn the credential.

Who This Is For

A good fit: engineers or architects whose work spans building, deploying, and governing agentic systems on Google Cloud. Google's recommended experience is 3+ years in industry building, testing, deploying, and managing cloud solutions, including 1+ year building agentic solutions on Google Cloud.

Not a fit: engineers who only write prompts. Part 1 of the exam (low-code) already tests whether you can assemble a runnable workflow with a workflow builder; Part 3 (custom agents) moves into coding. If you only want to prove you can apply an LLM, read PMLE instead.

How it relates to PMLE: PMLE had its exam guide rewritten in 2026 with agentic content absorbed in; PAA then promoted agents into their own certification. PMLE's skeleton is still classical ML engineering and it does not test multi-agent coordination; PAA section 3.3 does.

Official Specs at a Glance

ItemContent
Fee$200 (plus tax)
Duration2 hours
Questions and format50-60 multiple-choice questions
LanguageEnglish only
DeliveryOnline or onsite proctoring
PrereqsNone
Recommended experience3+ years plus 1 year on Google Cloud building agents
ValidityOne year

The exam assesses agentic workflow design rather than the model itself. The five capability areas Google names: building agents with low-code tools, using coding agents for application development, developing custom agents, evaluating and deploying agentic workflows, and securing and governing agentic workflows.

The exam is hybrid, two parts:

  • Part 1: proctored multiple-choice exam via Pearson, assessing conceptual knowledge, system design choices, architectural standards, cost optimization, and security governance.
  • Part 2: hands-on labs on Google Skills, unlocked through CM Connect after passing Part 1; both parts must be passed.

The Five Section Weights Are Your Study Plan

SectionWeightTopic
1. Building agents using low-code tools~13%Low-code agents
2. Using coding agents for application development~17%Coding agents
3. Developing custom agents~33%Custom agents
4. Evaluating and deploying agentic workflows~22%Evaluation and deployment
5. Securing and governing agentic workflows~15%Security and governance

The five add up to about 100%. Heaviest: section 3 (custom agents, 33%), then section 4 (evaluation and deployment, 22%).

Section 1: Low-code agents (~13%)

What it tests: assembling agentic workflows with Gemini Enterprise Workflow Builder and CX Agent Studio, including state-based workflows, system instructions, in-console prompt templates, and connecting enterprise data via Gemini Enterprise and Agent Search.

Read: official low-code tool docs and the official learning path.

Practice: build a runnable agent workflow with a workflow builder rather than only writing a system prompt.

Section 2: Coding agents (~17%)

What it tests: configuring coding agents (Antigravity, Claude Code on Google Cloud) with MCP servers, custom skills, and tools; using secure sandboxes such as GKE, Cloud Workstations, and Antigravity; refactoring, optimizing runtime, and patching application-layer vulnerabilities. Customization covers skills, plugins, extensions, hooks, rules, subagents, plus augmenting Antigravity with Agents CLI.

Read: Agents CLI and Antigravity docs.

Practice: bring a coding agent into an enterprise codebase workflow.

Section 3: Custom agents (~33%, heaviest)

What it tests: choosing between LLM and small language model, self-hosted vs SaaS, open-source vs proprietary; building with Agent Development Kit (ADK); sessions and memory; RAG with Vector Search and Agent Retrieval; agent permissions with Agent Identity and Agent Registry; orchestration with MCP, A2A, and multi-agent handoffs (parallel, sequential, graph).

Read: Agent Development Kit and Agent Platform docs.

Practice: run the full loop of custom agent, RAG, orchestration, handoff.

Section 4: Evaluation and deployment (~22%)

What it tests: test sets, golden data, continuous evaluation pipelines, eval tooling (ADK evalset, Agent Platform Gen AI evaluation service, custom autoraters); deployment runtimes (Agent Runtime, Cloud Run, GKE), troubleshooting, monitoring and optimization.

Practice: evaluate an agent, then deploy and monitor it.

Section 5: Security and governance (~15%)

What it tests: authentication, secure tool execution (OAuth 2.0), Principal Access Boundary (Agent Identity), Agent Gateway, Agent Registry, Model Armor; safety frameworks, guardrails, human-in-the-loop, secure data access.

A Schedule, and Why It Looks Like This

At 8-10 hours per week, plan roughly 8-10 weeks:

WeekContent
1Read the official exam guide, map to the official learning path
2Sections 1-2 (low-code agents, coding agents)
3-4Section 3 (custom agents, heaviest)
5-6Section 4 (evaluation and deployment)
7Section 5 (security and governance)
8Full review plus sample questions

Three Pitfalls

First, old study material has stale product names. Names like Workflow Builder, Agent Runtime, and Agent Search postdate much of the material in use during 2026; older guides still use Agent Designer, Agent Engine, and Vertex AI Search. Same trap as PMLE with Vertex AI.

Second, passing Part 1 is not the credential. Both parts are required. After GA results (7-10 business days), pass the labs within the 2-month window to receive the credential.

Third, respect the recommended experience. Without 1+ year of agentic building on Google Cloud, sections 3-5 will stall.

After the Exam: One-Year Validity and Continuing Education

PAA is valid for one year, not two like other Google professional certifications; Google cites the fast pace of the field and points to a continuing-education path to maintain active status.

What May Change (Re-check Before You Register)

ItemCurrent state (verified 2026-11-13)When to re-check
Certification statusGA (registration opened Nov 2, 2026)Quarterly
Exam structureHybrid two parts: Pearson MCQ + Google Skills labsExam guide updates
Duration and question count2 hours, 50-60 multiple-choice questionsEach exam guide update
Fee$200Quarterly
ValidityOne yearGoogle said it may expand
LanguageEnglishIf Chinese opens up
DeliveryOnline or onsite proctoring; results in 7-10 business daysPolicy page updates

Changelog

  • 2026-11-13: updated official specs from the GA page and FAQ (2 hours, 50-60 multiple-choice questions, $200, English, online or onsite proctoring, one-year validity); beta specs (Sept 8-30, 3 hours, ~80 questions, $120) moved to the comparison.

References

Site-internal related