Free Professional Agentic Architect practice test: 30 questions with answers

SequentialAgent executes its sub-agents in the listed order, which matches the requirement. ParallelAgent runs them concurrently, LoopAgent repeats them, and LlmAgent is a model-driven agent rather than an orchestration pattern.

Independent sources can be queried concurrently, so a parallel workflow shortens total wait time, and a final merge step combines the outputs. Sequential and single-agent designs wait on each call in turn, and a loop repeats work rather than gathering results from several sources.

A session is the record of one interaction, and long-term memory is a separate store that carries useful facts across sessions. New sessions do not inherit full histories by default, and deleting a session does not change the base model weights.

Reranking applies a more precise scoring model to a shortlist produced by first-pass retrieval. Chunking and embedding happen during indexing or the first retrieval step, and quantization changes how vectors are stored rather than how candidates are ranked.

A narrow, high-volume task that must stay in-house usually suits a small self-hosted model validated against the categories. Larger hosted models are not automatically more accurate, and the laptop and repeated-check designs fail the scale, cost or data-location requirements.

Agent Identity gives each agent its own identity so permissions can be scoped and audited per agent rather than shared. Persona settings, traffic inspection and public directories are different capabilities that do not establish who the agent is.

The evaluation set must cover the situations where the agent fails, so multi-turn and edge-case conversations drawn from real sessions reveal the gap. Adding easy single-turn items, removing failing cases or lowering the bar would make the score look better without measuring the problem.

An evalset packages evaluation cases, including the inputs and the expected behaviour, so an agent can be scored repeatably. The other options describe a serving endpoint, infrastructure provisioning and log export, which are not evaluation datasets.

A clear rubric and calibration against human labels make an autorater's judgments consistent and checkable. Removing the judge, raising its temperature or scoring fewer answers does not address the cause, which is an underspecified scoring standard.

A continuous evaluation pipeline runs the evaluation suite whenever something that affects agent behaviour changes, which surfaces regressions before they reach users. Automatic deployment, archival and response caching are separate concerns.

GKE provides cluster-level control over pods, sidecars and network policies, which matches the requirement. Agent Runtime is a managed agent service rather than a host for arbitrary clusters, and Cloud Run and Cloud Functions do not give the same cluster-level control.

Input data that shifts away from what the agent was tested on causes gradual accuracy loss even when the code is unchanged, so monitoring the inputs and refreshing the evaluation is the right response. The other causes would not produce a gradual decline tied to new product categories.

An isolated sandbox with short-lived, least-privilege credentials limits what a faulty or manipulated agent can reach. Keeping the production key, relying on a README note, or changing IDE settings does not contain the risk.

The Model Context Protocol is an open standard that lets agents discover and call tools and read resources or prompts from a server. Compilation, memory storage and IAM enforcement are handled by other tools or services, not by the protocol.

Rules provide always-on guidance, while a skill holds procedure-specific instructions that load on demand. Raising temperature adds randomness rather than consistency, and broad repository access does not define which standards the agent should follow.

Parameterized queries fix the root cause, and tests with malicious inputs plus human review confirm the fix before it merges. Bypassing review skips verification, a full rewrite is disproportionate, and disabling the endpoint leaves the flaw in the code.

A subagent works on a bounded task in a separate context and reports back, which keeps the main agent's context uncluttered. Delegation does not guarantee lower cost, bug-free output, or automatic commits, and tests are still needed.

Limiting the agent's MCP configuration to read and list tools removes the risky capability instead of relying on instructions. A prompt request can be ignored, a larger token budget does not change permissions, and moving credentials into a repository file exposes them.

User-delegated OAuth 2.0 with least-privilege scopes means the agent can only read what each employee is allowed to read. A shared admin key, pasted passwords or an unrestricted token all remove per-user limits and create a larger risk if the agent is misused.

Model Armor is a screening layer that inspects what goes into a model and what comes out for security and safety risks. Customer-managed encryption, audit retention and IAM assignment are real controls, but they are not what Model Armor does.

A principal access boundary policy sets the outer limit on the resources that principals can access, so roles granted elsewhere cannot exceed it. Token limits, prompt filtering and project creation are different controls that a boundary policy does not provide.

Keeping untrusted retrieved content separate from instructions and requiring human approval for sensitive actions are standard defenses against prompt injection. Broader mailbox access increases exposure, and letting the model judge injected instructions for itself is the weakness the attack relies on.

Propagating the end user's identity lets the backend apply per-user access checks, so the broad service account stops being the limit. Model refusals can be bypassed, date-based scoping still exposes recent orders across customers, and logging only reports leaks after they happen.

Agent Gateway is the capability for monitoring agent traffic and tracking agents. Model Armor screens content, a principal access boundary limits access and Agent Identity gives each agent its identity, so none of them provides traffic monitoring.

Pages with transition routes enforce a fixed order because a route fires only when its condition holds, so verification cannot be skipped. A single system prompt cannot guarantee the sequence, a separate agent per question is impractical, and re-reading history leaves the current step to guesswork rather than a defined state.

An event handler reacts to a named event, such as a timeout or unexpected input, with a defined response. Transition routes decide when the conversation changes page, few-shot examples shape output format, and IAM roles control access, so those are different mechanisms.

Connecting approved sources and requiring citations both tie each answer to company documents that users can verify. Relying on general knowledge can contradict the company's own policies, and a higher temperature adds variation in wording rather than factual grounding.

Multimodal models can work with images, audio and video when the originals are stored and indexed so relevant segments can be retrieved. Manual summaries are slow, file names discard the content, and a single frame loses most of what a video contains.

A permission-aware connector keeps each file's access rules in the search index, so users only see results they could open in the drive. Returning more documents, flattening group memberships or uploading files by hand does not enforce per-user access.

Targeted chain-of-thought with worked examples applies step-by-step reasoning where the multi-step logic sits, without cluttering simple lookups. Applying it everywhere adds latency and noise, deleting examples removes useful guidance, and temperature alone does not make a model reason.