Google Professional Agentic Architect: what the exam covers

The Google Cloud Certified Professional Agentic Architect is Google's certification for people who design, build and run AI agents on Google Cloud. It comes in 2 parts: a proctored multiple-choice exam (about 80 questions, 3 hours in the beta) and then a set of hands-on labs, and you are only certified once you pass both. It opens to everyone in mid-November 2026 at $200, and it lasts 1 year, half as long as Google's other professional certifications.

The exam guide is out, and it is more specific than most. A third of the exam is building agents in code. Nearly a fifth is about using coding agents, and it names Google's Antigravity and Claude Code on Google Cloud.

What the 5 sections test

Here are the sections in the guide's order, with what each one asks you to do in practice. Weights are Google's.

1. Building agents with low-code tools (~13%). State-based workflows of pages, transition routes and event handlers, in Gemini Enterprise Agent Designer and CX Agent Studio. System instructions and few-shot or chain-of-thought templates written in the console. Connecting Gemini Enterprise and Agent Search to company data, and bringing video, audio and images into a workflow.

2. Using coding agents for application development (~17%). Giving a coding agent MCP servers, custom skills and tools, with Antigravity and Claude Code on Google Cloud as the examples. Running coding agents in sandboxes (GKE, Cloud Workstations, Antigravity). Using them to refactor code, speed up runtimes and patch application vulnerabilities. Then customising them for a company: skills, plugins, extension hooks, rules and subagents in Antigravity, and Agents CLI for building and governing deployed agents. If you use Claude Code, our coding agents guide covers the same ideas: skills, hooks and subagents.

3. Developing custom agents (~33%). The biggest section, in 3 parts:

• Building in code: choosing a model (large or small, self-hosted or SaaS, open or proprietary) for cost, security and the agent's design; building with the open-source Agent Development Kit (ADK); configuring sessions and memory with Agent Platform Memory Bank.
• Company knowledge: RAG pipelines with embedding models, similarity scoring and reranking on Vector Search or Agent Retrieval; agent permissions through Agent Identity; Agent Registry and Google Cloud's MCP servers for connecting databases, APIs and outside SaaS tools.
• Orchestration: agents talking over MCP and Agent2Agent (A2A); handoffs between parallel, sequential and graph workflows with Agent Runtime and agent policies.

4. Evaluating and deploying agentic workflows (~22%). Building test sets from golden data, prompts and edge cases; continuous evaluation of an agent's tool calls against success criteria; choosing between ADK's evalset, the Gen AI evaluation service and custom autoraters. Then deployment: picking Agent Runtime, Cloud Run or GKE for the job and its cost, and troubleshooting drift, slow tool calls, reasoning loops and hallucinations. Our guide to evaluating an AI agent covers the method, whatever the cloud.

5. Securing and governing agentic workflows (~15%). OAuth 2.0 for agent-to-tool calls; principal access boundary policies on an agent's identity; Agent Gateway to watch traffic and track agents; Agent Registry and Model Armor for policy; guardrails and human-in-the-loop approval; carrying the user's identity through to the data.

Read the weights together and the exam is mostly about code: sections 2, 3 and 4 are 72% of it. The low-code section is the smallest.

Two parts: the exam, then the labs

The multiple-choice exam is delivered by Pearson, online or at a test centre, and tests "conceptual knowledge, system design choices, and architectural standards". Pass it and you get access to a separate set of labs in Google Skills, which test whether you can actually build what you chose. Passing the exam alone does not make you certified.

Two details matter for planning. The labs are not the ones in the learning path; they are a distinct set only for people who passed the exam. And in the beta they took under 5 hours, with 2 months to finish them. Google hasn't said whether the GA window is the same, so plan to do the labs soon after the exam, while the material is fresh.

The names have changed

Google renamed much of its agent stack in 2026, and the guide uses the new names. Older courses, blog posts and practice tests use the old ones:

The guide gives both renames itself, in its list of tools in scope. It also leans on names you won't find in older material: Agent Identity, Agent Registry, Agent Gateway, Skill Registry and Agents CLI. If a course doesn't cover them, it won't cover sections 3 and 5 the way the guide writes them.

The full list of tools in scope: ADK, agent evaluation, Agent Gateway, Agent Identity, Agent Registry, Agent Retrieval and Vector Search 1.0, Agent Runtime, Agent Search, A2A and MCP, Agents CLI, Antigravity (CLI, SDK and app), Auth Manager, BigQuery, Cloud Run, Cloud SQL, Cloud Storage, Firestore, Gemini Enterprise, Gemini models, Cloud Logging and Cloud Trace, GKE, Memorystore for Redis, Model Armor, MCP servers, Model Garden, RAG Engine, Sensitive Data Protection and Skill Registry.

Is it worth it, and is it for you?

Google recommends 3 or more years of building cloud solutions, including at least 1 year building agents on Google Cloud. There are no formal prerequisites, but that is a professional-level bar, and the labs mean you can't pass by memorising.

The 1-year validity is the trade-off. Google says it's because agent tools change so fast, and promises "a continuous education path" to keep the certification active, but hasn't said what that involves. Budget for keeping up every year, not for a 2-year credential.

It is the right exam if you build agents on Google Cloud. If you build on AWS, Azure or Databricks, the matching exam is in the best AI certifications in 2026. For Azure, AI-103 is the nearest equivalent, at associate level. If you build with Claude, there are Anthropic's own exams; our sample exams include the Claude Certified Architect and Developer.

How to prepare

1. Read the exam guide first and mark each consideration you haven't done by hand. It is 4 pages, and the list of tools in scope is the clearest view of what can come up.
2. Build one agent with ADK and run it on Agent Runtime. Give it memory, one RAG tool over Vector Search, one MCP tool, an agent identity and Model Armor in front, and turn on Cloud Trace. That one project touches most of sections 3, 4 and 5. Then deploy the same agent to Cloud Run, so you can say when each runtime fits.
3. Write an eval set and run it with ADK's evalset tooling, including a case where the agent should refuse or ask a person.
4. Use a coding agent for a week in a sandbox, with skills, hooks and an MCP server configured. Section 2 asks how to set these up for a team, not just how to prompt one.
5. Do Google's learning path (12 activities on Google Skills) and the sample questions on the certification page. Google calls the path "a starting point" and expects hands-on experience with every topic in the guide.

The GA exam replaces the beta, so its questions and pass mark may differ from what beta candidates saw. We'll update this page when it opens and Google publishes its final details.