Professional Agentic Architect study plan: four weeks, about 40 study hours

Where the 40 hours go

Hours follow the section weights in the exam guide, so the largest section gets the most time. If you already run ADK agents on Agent Runtime in production, skip the build steps you have done and keep the reading and the timed practice.

Before week 1, take the free 40-question practice exam once without studying. The score per section shows which week deserves extra time.

Week 1: build a custom agent with ADK (Section 3.1 and 3.2, about 10 hours)

• Build. Create an ADK agent with one function tool and run it in the ADK web UI (LLM agents, function tools). Store a value in session state and read it back on the next turn (state).
• Build. Add long-term memory with Memory Bank and confirm that a fact from one session is recalled in a new one (Memory Bank ADK quickstart).
• Build. Give the agent database access through MCP Toolbox for Databases with fixed, parameterized queries defined in tools.yaml (MCP Toolbox).
• Read. RAG Engine retrieval and reranking, hybrid search in Vector Search, and embedding task types (retrieval and ranking, hybrid search, task types).
• Read. Agent Identity: how a deployed agent gets its own principal and why a redeployment gets a new one (Agent Identity on Agent Runtime).
• Test. Section 3 practice. Read the explanation of every question you missed and open its documentation link.

Week 2: orchestration and security (Section 3.3 and Section 5, about 9 hours)

• Build. Turn two agents into a SequentialAgent, then a ParallelAgent, and note what each passes through session state (parallel agents, loop agents, graph workflows).
• Build. Expose one agent over A2A and call it from the other; open its Agent Card (A2A quickstart). Read where A2A ends and MCP begins (A2A and MCP).
• Build. Add a tool that requires human confirmation before it runs (tool confirmation).
• Read. Principal Access Boundary policies, 2-legged versus 3-legged OAuth for agents, Model Armor floor settings and Agent Gateway (PAB policies, 3-legged OAuth, floor settings, identity delegation).
• Test. Section 5 practice, then Section 3 again.

Week 3: evaluate and deploy (Section 4, about 9 hours)

• Build. Write an evalset for your week 1 agent and run adk eval with a tool trajectory criterion and a response criterion (ADK evaluation, criteria).
• Build. Deploy the same agent twice, once to Agent Runtime and once to Cloud Run, and compare what you had to configure (Agent Runtime, Cloud Run). Open its traces in Cloud Trace (tracing).
• Read. When to use the Gen AI evaluation service and online monitors instead of ADK evals, and how revisions and traffic splitting work on Agent Runtime (choose an evaluation, online monitors, revisions and traffic).
• Test. Section 4 practice.

Week 4: low-code and coding agents, then two simulations (Sections 1 and 2, about 12 hours plus 4 hours of simulations)

• Build. In Antigravity, add a workspace rule, a skill and a hook to a repository, and connect one MCP server through the workspace configuration so the whole team gets it (rules, skills, hooks, MCP). Run the Agents CLI lifecycle once (Agents CLI lifecycle).
• Read. Sandboxing coding agents on GKE and Cloud Workstations, and Claude Code on Google Cloud (GKE Agent Sandbox, Claude Code on Agent Platform).
• Read. Workflow Builder workflows, CX Agent Studio flows, handlers and handoff rules, and Gemini Enterprise data connectors (Workflow Builder, handlers, handoff, BigQuery connector).
• Test. Sections 1 and 2, then one full 50-question, 120-minute simulation in a single sitting. Use the weak-area drill on whatever scores lowest, and take a second simulation two days before the exam.

Pacing on exam day

At 2 hours for an estimated 50 to 60 questions, you have 2 to 2.4 minutes per question, and there are no breaks. In the simulations, flag any scenario you cannot settle within 3 minutes and come back to it after the first pass. The launch window, check-in and desk rules for online sessions are on the format and cost page.

After you pass: the labs

Passing the multiple-choice exam unlocks a separate set of labs in Google Skills, distinct from the labs in the learning path; you have 2 months to complete them. Beta candidates were told to expect two labs taking less than 5 hours in total. The build steps in weeks 1 to 4 are the closest preparation: the labs check that you can do this work in a live project. The official learning path (12 activities, to be updated at GA) adds guided labs on the same tools.