Google Cloud AI Certification Study Guide 2026: GenAI Leader, PMLE, Agentic Architect

By Kyle Tice · Updated October 19, 2026 · Facts checked against official exam pages

Take the Generative AI Leader first only if you’re in a business or strategy role. Engineers should go straight to the Professional Machine Learning Engineer (PMLE), then the new Professional Agentic Architect once it reaches general availability in mid-November 2026. If agents are your job and ML training isn’t, you can skip PMLE and go directly to Agentic Architect.

This guide covers the 2026 changes, official section weights, week-by-week plans, resources and the labs that matter.

What changed in 2026

• PMLE got a new exam guide dated June 1, 2026. Google’s page says the exam now reflects the move from Vertex AI to the Gemini Enterprise Agent Platform, updates to the data and analytics stack, and more emphasis on Google Cloud-native solutions. Product names on the exam use the new branding (Agent Platform AutoML, Agent Platform Pipelines, Agent Platform Feature Store, Model Garden). Older courses that say “Vertex AI” still teach the right concepts, but you need to map the names.
• The Professional Agentic Architect is new. It’s Google’s first certification that requires both a multiple-choice exam and hands-on labs. The beta ran in September 2026 at $120. Google says GA registration opens November 2 and the GA exam launches in mid-November 2026 at $200.
• Agentic Architect is valid for only one year, compared with two years for other Google Cloud professional certifications. Google offers a continuing education path to keep it active.
• Generative AI Leader was updated for branding. The exam guide now refers to Gemini Enterprise, Agent Search (formerly Vertex AI Search) and Agent Platform.

Which Google Cloud AI exam to take, and in what order

As of October 2026: Generative AI Leader costs $99 (90 minutes, 50–60 questions, valid three years). PMLE costs $200 (two hours, 50–60 questions). Agentic Architect costs $200 at GA. None has a formal prerequisite.

Who should skip: Engineers get little from the Generative AI Leader; it tests business-level product knowledge, not implementation. And if your company doesn’t run on Google Cloud, a public agent project on GitHub will usually beat a GCP-specific credential.

Generative AI Leader

Exam sections (official)

Section 2 is the biggest and the most Google-specific: Gemini app, Gemini Enterprise, Gemini for Google Workspace, Customer Engagement Suite, Agent Platform, Model Garden, RAG offerings and agent tooling. Section 3 covers grounding, RAG, prompt techniques (zero-shot through ReAct) and sampling parameters. Section 4 covers Google’s Secure AI Framework (SAIF) and responsible AI.

2–3 week plan (about 15–30 hours)

• Week 1: Fundamentals. ML approaches, the ML lifecycle, foundation model selection factors, data quality, Gemini vs. Gemma vs. Imagen vs. Veo.
• Week 2: Google’s product map. Make a one-page table: product, who uses it, what problem it solves. This section is 35% of the exam and mostly tests whether you can match a scenario to a product.
• Week 3: Grounding, prompt techniques, SAIF and responsible AI. Take the sample questions and book.

Professional Machine Learning Engineer

Exam sections (official, as of June 1, 2026)

Google notes the exam doesn’t directly assess coding, but you should read Python and SQL. Recommended experience is three or more years in industry, including a year designing and managing solutions on Google Cloud.

New or more prominent in this revision: fine-tuning Gemini models with BigQuery, evaluating gen AI with LLM-as-a-judge, Gemini cost and latency optimization, Model Armor for prompt and data security, and monitoring gen AI solutions alongside classic drift and skew.

10-week plan (about 80–120 hours)

• Weeks 1–2: Low-code. BigQuery ML (classification, regression, forecasting, clustering, feature engineering, predictions), Agent Platform AutoML, Model Garden model selection, Document AI and Vision APIs.
• Weeks 3–4: Data and experimentation. Choosing BigQuery vs. Dataflow vs. Spark, Feature Store, Workbench and Colab Enterprise, experiment tracking and ML Metadata, evaluation metrics.
• Weeks 5–6: Training at scale. Custom training, Kubeflow on GKE, hyperparameter tuning, fine-tuning foundation models, CPU vs. GPU vs. TPU, data and model parallelism.
• Week 7: Serving. Batch vs. online, prebuilt and custom containers, Model Registry, A/B and canary rollouts, public vs. private endpoints, autoscaling.
• Week 8: Pipelines. Agent Platform Pipelines, Managed Service for Apache Airflow, Ray, CI/CD/CT with Cloud Build, retraining policies.
• Week 9: Monitoring and security. Model Monitoring, training-serving skew, data and concept drift, explainability, Model Armor, bias monitoring.
• Week 10: Practice exams and review.

Professional Agentic Architect

Exam sections (official)

The in-scope tool list includes the Agent Development Kit (ADK), Agent Runtime (formerly Agent Engine), Agent Search, Agent Identity, Agent Registry, Agent Gateway, Agents CLI, Antigravity, Google Cloud MCP servers, RAG Engine, Vector Search, Model Armor, Gemini Enterprise, Cloud Run and GKE. Section 2 also names Claude Code on Google Cloud as an example coding agent.

How the exam and labs work

1. Part 1: a proctored multiple-choice exam through Pearson (online or onsite). The beta was three hours and about 80 questions; Google’s FAQ says the GA exam is two hours.
2. Part 2: hands-on labs on Google Skills, unlocked only after you pass Part 1. Google’s FAQ says you have two months after passing the exam to finish the labs.
3. You must pass both to earn the certification.

Google strongly recommends three or more years building cloud solutions plus at least one year building agentic solutions on Google Cloud.

10–12 week plan (about 80–120 hours)

• Weeks 1–2: Low-code agents. Gemini Enterprise Agent Designer, CX Agent Studio state-based flows (pages, routes, event handlers), connecting enterprise data through Agent Search.
• Week 3: Coding agents. Configure a coding agent with MCP servers and custom skills, run it in a sandbox (Cloud Workstations or GKE), and use it to refactor code. Learn Antigravity’s skills, hooks, rules and subagents.
• Weeks 4–6: Custom agents with ADK, the biggest section. Model selection (LLM vs. SLM, hosted vs. self-hosted), sessions and Memory Bank, RAG with embeddings and reranking, Agent Identity permissions, MCP and A2A, sequential, parallel and graph multi-agent workflows.
• Weeks 7–8: Evaluation and deployment. Build golden test sets, use ADK evalsets and the Gen AI evaluation service, pick between Agent Runtime, Cloud Run and GKE, troubleshoot reasoning loops and tool latency.
• Week 9: Security and governance. OAuth 2.0 for agent-to-tool calls, principal access boundary policies, Agent Gateway, Model Armor, human-in-the-loop.
• Weeks 10–12: Practice questions, then rehearse lab-style tasks under time pressure.

Free official resources first

1. Exam guides (linked in Sources). Print the section lists and use them as a checklist.
2. Google Skills (formerly Google Cloud Skills Boost). Each exam has a learning path. The Agentic Architect path is linked from the official cert page. Most labs run in temporary projects, so you don’t need your own billing account for them.
3. Sample questions on each official cert page.
4. Kaggle’s 5-Day AI Agents Intensive (self-paced) for agent fundamentals with Gemini and ADK.
5. Google Cloud docs for ADK, Agent Platform and BigQuery ML. The docs are where renamed products are explained.

Paid resources

• Coursera hosts Google Cloud’s official specializations, including ML engineering tracks. If you like structured courses with graded labs, start here. See Google Cloud courses on Coursera.
• Official PMLE study guide book (Wiley), referenced on Google’s PMLE page. Check that the edition you buy is current; older editions use Vertex AI names.
• Practice tests. Whizlabs and Udemy sellers publish PMLE practice exams. Look for ones updated after June 2026 that use Agent Platform naming. Browse Whizlabs Google Cloud practice tests or search PMLE courses on Udemy.

Third-party material for Agentic Architect will be thin until after GA. For now, the exam guide, Google Skills path and ADK docs are your main sources.

Practice-exam strategy

• Google’s questions are long scenarios. Most give you a company, constraints and four plausible answers. Practice reading for the constraint that rules out three of them: “minimal code”, “lowest latency”, “managed service”, “data must stay in region”.
• Prefer managed services unless the question says otherwise. Google’s guides emphasize Google Cloud-native solutions.
• Map old names to new ones. If a practice test says “Vertex AI Pipelines”, translate it to Agent Platform Pipelines. The real exam uses the new names.
• For Agentic Architect, practice building, not just answering. The labs are a separate pass/fail gate. Time yourself on lab-style tasks.
• No dumps. They violate Google’s certification terms.

Hands-on labs to do

Generative AI Leader: none are required, but spend an hour in the Gemini app and Google AI Studio, and walk through one Agent Search demo so the product names mean something.

PMLE

1. Train a BigQuery ML model, evaluate it and generate predictions with SQL only.
2. Run a custom training job, register the model, deploy it to an endpoint and split traffic between two versions.
3. Build a pipeline with data validation, training, evaluation and conditional deployment.
4. Turn on Model Monitoring and simulate drift.
5. Fine-tune a Gemini model on a small dataset and evaluate it against the base model with LLM-as-a-judge.

Agentic Architect

1. Build an ADK agent with two tools and session memory, and deploy it to Agent Runtime.
2. Add a RAG tool over your own documents with Vector Search and reranking.
3. Connect the agent to an MCP server and a second agent over A2A.
4. Write an evalset of 20–30 cases and run it after every change.
5. Put Model Armor in front of the agent and test it with a prompt-injection attempt.

These double as portfolio projects. See /guides/ai-portfolio-projects/.

Related reading

• AIF-C01 vs Google Generative AI Leader
• AWS MLA vs Google PMLE
• Google Agentic Architect vs NVIDIA Agentic AI
• Vouchers, including Google’s GEAR Get Certified program

FAQ

Can I still use Vertex AI study material for PMLE?

Partly. The concepts (pipelines, feature stores, monitoring, tuning) carry over. But the June 2026 guide uses Gemini Enterprise Agent Platform names and adds gen AI evaluation, Gemini fine-tuning and Model Armor. Use the new exam guide as your checklist.

When can I take the Agentic Architect exam?

Google says GA registration opens November 2, 2026 and the exam launches in mid-November 2026. Check the official page before you book, since launch dates can move.

How do I keep Agentic Architect active?

It’s valid for one year, half the usual two years for Google Cloud professional certs. Google’s FAQ says a continuing education path is available to keep it active. Budget time for that each year, and expect the content to change as agent tooling does.

Do I need PMLE before Agentic Architect?

No. There are no formal prerequisites. PMLE helps if you’ll be tuning or serving models, but the Agentic Architect exam focuses on building, evaluating and securing agents.

Is the Generative AI Leader worth it for engineers?

Usually not. It’s a business-level exam. Engineers get more from PMLE or Agentic Architect plus a public project.