What You Need to Know to Pass the Professional Agentic Architect Exam

GCP Study Hub
November 15, 2026

The Professional Agentic Architect exam tests designing, building, deploying, grounding, evaluating, observing, securing, and governing AI agents on Google Cloud. An agent is a model given instructions and tools that runs in a loop, deciding whether to call a tool, reading the result, and repeating until it can answer. Agents can also be composed into workflows that route requests, run work in parallel or sequence, loop until output meets criteria, or combine agents with plain code in a graph.

The exam spans custom-built agents (ADK, Agent Runtime, MCP, and A2A) alongside managed and coding products (Gemini Enterprise, CX Agent Studio, Antigravity, Claude Code on Agent Platform, and CodeMender). Across these custom and managed options, preparation centers on how the Google Cloud agent stack fits together and where each architectural boundary sits.

How the Google Cloud agent stack fits together

The exam topics group into four broad areas:

  • Models and protocols: Gemini on Agent Platform, model routing, quotas, structured output, Model Garden (Model as a Service versus self-deployment), self-hosting open models on GKE, MCP, OpenAPI, and A2A.
  • Building, state, and deployment: ADK tools, workflows, callbacks, and tool confirmation, Agent Platform Sessions and Memory Bank, and deployment on Agent Runtime, Cloud Run, or GKE.
  • Grounding, evaluation, and observability: Agent Search, RAG Engine, Vector Search, BigQuery embeddings, response and trajectory evaluation, and OpenTelemetry tracing.
  • Security, governance, and builders: Agent Identity, Agent Gateway, Agent Registry, Model Armor, Gemini Enterprise, CX Agent Studio, and coding agents such as Antigravity.

Foundational architectural boundaries

Three recurring boundaries appear across the Professional Agentic Architect exam.

The first is tools versus agents. A plain operation with no reasoning and no multi-step conversation, such as looking up an order or creating a return label, belongs as a tool. Wrapping it in an agent adds latency and network hops without adding value. Use an agent when a capability needs its own reasoning, owns its own models and tools, or belongs to an external party that does not share how it works, such as a supplier's quoting agent.

The second is guidance versus enforcement. Prompt instructions, docstrings, and rule files are guidance that a model usually follows without guarantee. When a rule must be enforced or an action must never occur, place the control outside the model: in code (callbacks, argument validation, tool confirmation, or graph routing), in configuration (policies, hooks, sandboxes, or parameterized queries), or in managed services (Model Armor, Agent Gateway, or Agent Search access control). When enforcement must sit outside the agent's own code, external controls such as gateway policies replace in-code filters.

The third is deterministic versus probabilistic controls. Check fixed formats with code or regular expressions, and reserve model-based judges (including classifier models such as Llama Guard) for policies about meaning or phrasing. Lowering temperature makes outputs more repeatable rather than more correct, and a response schema provides structure guarantees that a prompt instruction cannot.

Connecting tools and agents: MCP, A2A, and sub-agents

Communication patterns follow ownership and coupling:

  • MCP connects an agent to its own tools and data, including shared tool servers and databases, alongside OpenAPI specifications and enterprise connectors.
  • A2A connects to an independent agent owned by another team or company, or shared across multiple applications.
  • Local sub-agents in the same ADK application serve tightly coupled agents owned by the same team, running in one process without network calls or A2A serialization overhead.

The deciding question is who else needs the agent. Folding a shared agent into a local sub-agent cuts off its other callers and owning team, keeping an unshared agent behind A2A adds unnecessary latency, and wrapping another team's agent as MCP tools treats a peer agent as a toolbox.

Runtime choices, sessions, and memory

Google's deployment guidance follows an operational overhead ladder, choosing the most managed option that meets requirements and stepping down only when a constraint forces it, such as regulatory self-hosting. When workloads have different constraints, split them and self-host only the constrained part.

  • Agent Runtime is the managed platform for Python agents with minimal operations, bundling Sessions, Memory Bank, autoscaling, and built-in logging, monitoring, and tracing. It is built for Python agents and does not host custom MCP servers.
  • Cloud Run fits containerized workloads needing serverless scale-to-zero and language flexibility, such as a TypeScript ADK agent or a custom MCP server over streamable HTTP. Because Cloud Run is stateless, state belongs in Agent Platform Sessions or Firestore.
  • GKE (Standard or Autopilot) fits complex stateful workloads, fine-grained infrastructure control, sandboxed code execution, and self-hosted models when the team can operate Kubernetes.
  • Compute Engine managed instance groups leave scaling, patching, sessions, memory, and tracing to the team.

State divides between sessions and long-term memory. A session holds one conversation's event history and state. In-memory sessions vanish on restarts or deployments, and database storage requires operating a database, whereas Agent Platform Sessions persists conversations with no database to run, even behind Cloud Run. Memory Bank is managed long-term memory per user that extracts facts across conversations and consolidates updates or contradictions when preferences change. It is not a session store, cross-user cache, grounding source, or tracker for changing task status.

Grounding, evaluation, and observability

Grounding ties answers to trusted sources to reduce hallucinations, and it is the right fix when an agent invents unsupported details rather than lowering temperature. ADK supports grounding against internal Agent Search data stores (returning grounding metadata for used chunks) and public web search, which cannot access internal content or fix reasoning errors when facts are already in the prompt. Grounding also spans RAG Engine, Vector Search filtering with restricts, and BigQuery embeddings.

Evaluation checks both the final response and the trajectory. Final response metrics measure answer quality, while trajectory metrics verify which tools the agent called, with which arguments, and in what order. An agent can answer correctly while taking a flawed tool path that unit tests and traces alone do not catch, because traces show timing rather than correctness. For observability, the exam covers OpenTelemetry in ADK, agent telemetry on Agent Runtime, metrics and alerts, reading traces, capturing prompts and responses, and distinguishing slow agents from wrong ones.

Identity, traffic governance, and Model Armor

Shared user-managed service accounts with keys create three problems: agents look identical in audit logs, long-lived keys can leak, and anyone with the Service Account Token Creator role can impersonate the account. Agent Identity gives each agent its own SPIFFE-based cryptographic identity that appears individually in Cloud Audit Logs, receives IAM roles directly, cannot be impersonated, and uses access tokens bound to X.509 certificates renewed every 24 hours without keys. Service-account policies do not apply to these agents.

Traffic governance depends on direction, because a control on the wrong side never inspects the traffic. Bound to an agent with Agent Identity on Agent Runtime, Agent Gateway enforces policies centrally outside agent code:

  • Client-to-Agent governs inbound user traffic, screening prompts with a regional Model Armor template in the same region as the gateway (templates rather than floor settings).
  • Agent-to-Anywhere governs outbound traffic by checking destinations against Agent Registry (the catalog of approved agents, MCP servers, tools, and external REST endpoints) with default-deny egress policies. IP-based controls such as Cloud NAT or firewall rules cannot govern MCP or A2A traffic.

Model Armor screens prompts and responses for prompt injection and jailbreaks, responsible AI categories, malicious URLs, and sensitive data through Sensitive Data Protection using inspect-and-block or inspect-only enforcement. Across security controls, match each service to its scope: VPC Service Controls restricts where data can move, principal access boundaries limit what principals can reach, IAM governs Google Cloud resource permissions (not individual MCP tools), organization policies govern resource configuration, Agent Gateway governs which tools, APIs, and agents an agent may call, and search-level access control restricts which documents a user may see.

Choosing between Gemini Enterprise, CX Agent Studio, ADK, and coding agents

Builder choice follows who the agent serves and who maintains it:

  • Gemini Enterprise and Workflow Builder serve a company's own employees with agents and AI search connected to internal sources such as Microsoft SharePoint, Gmail, and Google Drive. Non-developers can manage employee workflows, including scheduled tasks, and ADK agents on Agent Runtime can be published there for employees without exposing them as tools to other agents or external companies.
  • CX Agent Studio, part of Gemini Enterprise for Customer Experience, builds customer-facing agents across telephony, voice, and web chat. As the evolution of Dialogflow CX (while Dialogflow ES is legacy), it gives analysts a low-code builder with flow reuse, human handoff rules, evaluations, and traffic splitting.
  • ADK on Agent Runtime or Cloud Run fits custom agents written and maintained in code by developers.

The exam also covers coding agents, centered on Antigravity (desktop app and IDE, CLI, JetBrains IDE extension, and Python SDK customized with rules, skills, subagents, hooks, MCP servers, and plugins), alongside Claude Code on Agent Platform, CodeMender, and Cloud Workstations.

How to approach exam preparation

When preparing for the Professional Agentic Architect exam, we would focus on the architectural boundaries first: tools versus agents, guidance versus enforcement, deterministic versus model-based checks, inbound versus outbound traffic, and sessions versus long-term memory. Then map requirements along the operational overhead ladder and distinguish among Google Cloud's access controls, governance services, and builders.

Our Professional Agentic Architect practice exams test the full Google Cloud agent stack alongside ADK architecture, evaluation, and governance controls, with detailed explanations that drill these distinctions.

Get tips and updates from GCP Study Hub

arrow