Google Professional Agentic Architect: what it is

Professional Agentic Architect is Google Cloud's new certification for developers and architects who design, build and run AI agents on Google Cloud. As of October 21, 2026 it is in beta, with registration closed. Google says the generally available exam launches in mid-November 2026, at 200 USD for a two-hour exam followed by hands-on labs.

In Google Trends over the past 90 days, "professional agentic architect" was a Breakout rising query among searches for "google cloud certification", and some pages about it still quote the beta figures. This guide covers what the certification is, the beta and launch timeline, its unusual two-part format, the five exam domains and the tools behind them, who should take it, and how to prepare. Every fact was checked on Google Cloud's certification page, exam guide and FAQ on October 21, 2026.

Professional Agentic Architect at a glance

Full name: Google Cloud Certified – Professional Agentic Architect. Like all Google Cloud exams, it has no exam code.
Status (October 2026): beta. Registration is closed, and general availability is due in mid-November 2026.
Who it's for: experienced developers or architects who design and manage autonomous agentic workflows on Google Cloud.
Recommended experience: three or more years building and deploying cloud solutions, including at least one year building agentic solutions on Google Cloud. There's no formal prerequisite.
Format: two parts. Part 1 is a proctored multiple-choice exam delivered by Pearson, online or at a test centre. Part 2 is a set of hands-on labs in Google Skills. You must pass both.
Price: 200 USD at general availability. The beta cost 120 USD plus tax.
Language: English for the beta. Google hasn't listed launch languages.
Passing score: not published.
Validity: one year.

The official details are on the Professional Agentic Architect page and in the exam guide.

Beta and general availability: the timeline

Google ran the beta in September 2026. Registration opened on September 3 and closed on September 30, and the beta exam window ran from September 8 to September 30. Google says beta results come out at the end of October, and beta candidates who pass Part 1 have until December 31, 2026 to finish the labs.

The generally available exam differs from the beta, so check which figures a page is quoting:

Beta: about 80 multiple-choice questions in three hours, for 120 USD (a 40% discount).
General availability: a two-hour session for 200 USD, with results in 7 to 10 business days. Google hasn't published the question count.

If you're reading this before mid-November 2026, you can't book yet. Google's English page says only "mid-November", so watch the certification page for the registration opening.

How the two-part format works

Most Google Cloud certifications are a single multiple-choice exam. Professional Agentic Architect adds a practical second part, unlike Generative AI Leader or Professional Machine Learning Engineer.

Part 1, the exam. A proctored multiple-choice exam that tests concepts and design decisions.
Part 2, the labs. Passing Part 1 unlocks hands-on labs in Google Skills, through your certification account. Google's FAQ says the labs take under five hours, and you have two months to complete them once they're available.

Only when both parts are passed do you hold the certification. Google hasn't published how many labs there are or how they're scored, so plan to be comfortable building, not just choosing, before you book.

The five domains in practice

The exam guide splits the content into five domains. Custom agents is the largest, at about a third.

Building agents with low-code tools (about 13%)

This domain covers agents built without much code: Gemini Enterprise's Agent Designer (now called Workflow Builder in Google's docs) and CX Agent Studio, connecting them to enterprise data through Agent Search (formerly Vertex AI Search), and handling multimodal data. A related question is when a no-code agent is enough, and this practice question on a no-code custom agent asks it from a business user's side.

Working with coding agents (about 17%)

This is the newest material. It covers giving coding agents MCP servers, skills and tools, running them in sandboxes such as GKE or Cloud Workstations, and using them to refactor and patch code. The guide names Google's Antigravity (its skills, rules and subagents) and the Agents CLI, and also names Claude Code on Google Cloud as an example of a coding agent.

Building custom agents (about 33%)

A third of the exam is building agents with code. That means:

choosing a model: a large or small language model, open-source or proprietary;
building with the Agent Development Kit (ADK), including sessions and Memory Bank for state;
grounding agents with retrieval-augmented generation and vector retrieval;
giving agents an identity and registering them in the Agent Registry;
orchestrating several agents with MCP and the Agent2Agent (A2A) protocol in parallel, sequential and graph workflows, and running them on Agent Runtime (formerly Agent Engine).

The core idea behind an agent is a loop of reasoning, acting with a tool and observing the result, which this practice question on ReAct prompting covers. Agents need tools because grounding alone can't change a booking or look up live status, as this practice question on giving an assistant access to live systems shows. If retrieval is new to you, start with the guide to retrieval-augmented generation.

Evaluating and deploying agents (about 22%)

This domain covers proving an agent works and keeping it working: golden test sets, continuous evaluation, ADK evaluation sets, Google's Gen AI evaluation service and custom autoraters. On the deployment side, it covers choosing a runtime (Agent Runtime, Cloud Run or GKE), troubleshooting drift, latency and reasoning loops, and monitoring cost and quality.

Securing and governing agents (about 15%)

The last domain covers authorising tool calls with OAuth 2.0, Agent Identity and its policies, Agent Gateway, the Agent Registry, Model Armor, and keeping a human in the loop for risky actions. Model Armor also appears in Google's Machine Learning Engineer exam, as in this practice question on screening prompts with Model Armor.

The tools and protocols to know

The exam guide lists the products in scope. Grouped:

Agent building: Agent Development Kit, Agents CLI, Antigravity (CLI, SDK and app), Gemini models, Model Garden and Gemini Enterprise.
Agent platform: Agent Runtime, Agent Search, Agent Registry, Agent Identity, Agent Gateway, agent evaluation, RAG Engine, Agent Retrieval and Vector Search, and the Skill Registry.
Protocols: MCP (Model Context Protocol) and A2A (Agent2Agent), plus Google's own MCP servers.
Security: Auth Manager for OAuth 2.0, Model Armor and Sensitive Data Protection.
Supporting services: BigQuery, Cloud Run, Cloud SQL, Cloud Storage, Firestore, GKE, Memorystore for Redis, and Cloud Logging and Trace.

Two renames matter for study material: Agent Runtime used to be Agent Engine, and Agent Search used to be Vertex AI Search. Material that uses only the old names may be out of date.

For a map of Google's wider AI product family, see the guide to Google Cloud's generative AI offerings.

Who should take it

Professional Agentic Architect fits people already building agents on Google Cloud: developers wiring agents to company systems, architects designing multi-agent workflows, and platform engineers who deploy and secure them. Google recommends at least a year of agentic work on Google Cloud, and the lab part means you can't pass on reading alone.

It's less suitable if:

You're new to generative AI. Generative AI Leader (99 USD, valid three years) is the non-technical starting point, and the Generative AI Leader practice tests cover it.
You train and serve models rather than build agents. Professional Machine Learning Engineer is the closer fit. Practice for it with the Machine Learning Engineer practice tests.
You need the credential to last. It's valid for one year, and Google hasn't yet published how renewal will work.

Google's pages don't describe a ladder between these certifications. None is a prerequisite for another, so you can go straight to Professional Agentic Architect if your experience fits.

How it differs from Google's other AI certifications

Audience: Generative AI Leader is for business and non-technical roles. Machine Learning Engineer is for people who build and run models. Agentic Architect is for people who build and run agents.
Format: Generative AI Leader and Machine Learning Engineer are single multiple-choice exams. Agentic Architect is an exam plus hands-on labs.
Price (as of October 2026): Generative AI Leader costs 99 USD. Machine Learning Engineer and Agentic Architect cost 200 USD each.
Validity: Generative AI Leader lasts three years. Agentic Architect lasts one year.
Status: Generative AI Leader and Machine Learning Engineer are generally available. Agentic Architect is in beta until mid-November 2026.

If you're comparing across vendors, the best AI certifications guide puts Google's AI exams next to AWS, Microsoft and others.

Cost, validity and preparation

As of October 2026, the certification costs 200 USD once generally available. Google's pages list no rupee price. It's valid for one year. Google's FAQ says a "continuous education path" will let you keep it active, but the details aren't published yet.

Google's preparation resources:

The Professional Agentic Architect learning path on Google Skills (path 4525). Google calls it a starting point that will be updated at launch.
The exam guide, with the full list of skills and tools.
Sample questions, through a form linked from the certification page.
Skill badges on Google Skills for agent topics.

There's no Professional Agentic Architect practice bank on this site. The closest practice is the agent, grounding and safety material in the Generative AI Leader question list, such as this practice question on moving a prototype to an enterprise platform.

A practical plan: build one agent with ADK that uses an MCP tool and a RAG source, deploy it to Agent Runtime, add an evaluation set and an OAuth-protected tool, and then read the exam guide line by line for anything you haven't touched. The labs will expect that kind of fluency.

Frequently asked questions

Is the Professional Agentic Architect exam available now?

Not yet. As of October 21, 2026 it's in beta with registration closed. Google says the generally available exam launches in mid-November 2026.

How much does the Professional Agentic Architect certification cost?

200 USD at general availability. The beta cost 120 USD, a 40% discount, and that offer has closed.

How long is it valid?

One year, according to Google's certification page and FAQ. That's shorter than most Google Cloud certifications.

Does passing the exam give me the certification?

No. You must pass Part 1, the multiple-choice exam, and then Part 2, the hands-on labs in Google Skills, within two months of the labs becoming available to you.

Do I need Generative AI Leader or another certification first?

No. Google lists no prerequisites. It recommends three or more years of cloud experience, including a year building agentic solutions on Google Cloud.

How hard is it?

Google hasn't published a passing score or pass rate. The format makes it demanding: a professional exam followed by labs, covering a set of products that were renamed or launched recently.