Google Cloud Professional Agentic Architect

Learning plan

About 100 hours of focused study, split into 6 phases. Pick your weekly pace and the timeline adjusts.

Estimated finish: 10 weeks

What the exam covers

Source: official exam guide. Weights can change, so check before you start.

Before you start

• A Google Cloud project with billing enabled (new accounts get $300 in free credits) and Gemini Enterprise access for the low-code section
• Fluent Python, plus REST APIs, OAuth 2.0 and containers (Cloud Run or GKE)
• Google recommends 3+ years building cloud solutions and 1+ year building agents on Google Cloud

1. Weeks 1–2 · 12 h

   Low-code agents and enterprise data

   • Build a state-based agent in CX Agent Studio with pages, transition routes and event handlers, and a task agent in Gemini Enterprise Agent Designer
   • Write system instructions and few-shot or chain-of-thought prompt templates in the console
   • Connect Gemini Enterprise and Agent Search (formerly Vertex AI Search) to enterprise data sources, including video, audio and image content

   Use: Professional Agentic Architect learning path (Google Skills) · Agentic Architect exam guide (PDF)

2. Weeks 2–3 · 14 h

   Coding agents for development

   • Configure Antigravity and Claude Code on Google Cloud with MCP servers, custom skills and tool access
   • Run coding agents in sandboxes (GKE, Cloud Workstations, Antigravity) and use them to refactor code and patch vulnerabilities
   • Create skills, plugins, hooks, rules and subagents in Antigravity, and extend it with Agents CLI to deploy and govern agents

   Hands-on: In a Cloud Workstation, connect a coding agent to a Google Cloud MCP server (for example BigQuery), add one custom skill and one pre-commit hook, and have the agent fix a seeded SQL-injection bug in a small Flask app.

   Use: Model Context Protocol docs

3. Weeks 3–6 · 32 h

   Custom agents with ADK (largest domain)

   • Build agents with the Agent Development Kit (LlmAgent, SequentialAgent, ParallelAgent, LoopAgent, graph workflows) and pick LLM vs SLM, hosted vs self-hosted
   • Add sessions and long-term memory with managed sessions and Agent Platform Memory Bank
   • Build RAG with RAG Engine or Vector Search / Agent Retrieval, including embedding choice, similarity scoring and reranking
   • Connect agents to tools through Google Cloud MCP Servers and Agent Registry, and to other agents through Agent2Agent (A2A)
   • Scope each agent's permissions with Agent Identity

   Hands-on: With ADK, build a coordinator agent that delegates to two sub-agents (a RAG agent over Cloud Storage documents via RAG Engine, and a tool agent that calls a Cloud Run API through MCP). Add Memory Bank so it remembers user preferences between sessions. Expose one sub-agent over A2A.

   Use: Agent Development Kit docs · ADK samples (GitHub) · Agent2Agent (A2A) protocol · Kaggle 5-Day AI Agents Intensive (self-paced)

4. Weeks 6–8 · 22 h

   Evaluation, deployment and operations

   • Write golden datasets and ADK evalsets that test both final responses and tool-call trajectories, and add custom autoraters with the Gen AI evaluation service
   • Choose Agent Runtime (formerly Agent Engine), Cloud Run or GKE by scaling, cost and control needs
   • Trace agents with Cloud Trace and Cloud Logging, and fix reasoning loops, tool-call latency, drift and hallucinations

   Hands-on: Write a 15-case evalset for your ADK agent, run `adk eval`, deploy the agent to Agent Runtime, and use Cloud Trace to find its slowest tool call. Then change one prompt and rerun the evalset to catch regressions.

   Use: Agent Starter Pack (GitHub) · Gen AI evaluation service overview

5. Weeks 9–10 · 12 h

   Security and governance

   • Secure agent-to-tool calls with OAuth 2.0 through Auth Manager, and set principal access boundary (PAB) policies with Agent Identity
   • Monitor and control agent traffic with Agent Gateway, and enforce policy through Agent Registry and Model Armor
   • Design human-in-the-loop checkpoints and propagate end-user identity to data sources

   Use: Model Armor

6. Week 10 · 8 h

   Practice and lab readiness

   • Review every in-scope tool listed on page 4 of the exam guide and write one sentence on when you'd choose it
   • Rebuild the ADK lab from scratch without notes in under 3 hours, because the hands-on labs test real execution
   • Do timed 2-hour sets of scenario questions on architecture trade-offs

   Use: Agentic Architect certification details (Google support)

You're ready to book when…

• You can build, evaluate and deploy a multi-agent ADK app to Agent Runtime from a blank project without following a tutorial
• You can explain when to use MCP vs A2A, and Agent Identity vs Agent Gateway vs Model Armor
• You can pick Agent Runtime vs Cloud Run vs GKE for a given scale, cost and control requirement
• Your ADK evalset passes consistently after a prompt or model change

Exam tips

• You need to pass both parts: the multiple-choice exam (Pearson) and hands-on labs (Google Skills). The labs unlock only after you pass the exam, and you then have 2 months to finish them.
• The credential is valid for 1 year, not 2, so plan for the continuing-education path to keep it active.
• Custom agents carry ~33% of the exam, so ADK fluency is worth more than any other single topic.