Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 17

Question 160

A retail agent must remember an item during one chat. Which agent/tool architecture control directly resolves the situation?

Store the shopping-cart state in a shared agent-global variable accessible to all sessions.
Write every one-turn shopping item permanently into all customers' long-term memory.
Rely on the model's pretrained weights to remember the item on the next API call.
Keep short-term session state isolated to the current authenticated actor.

Explanation

Correct Answer

Use actor-bound session state for the current chat; short-term context must not bleed between user or session identifiers.

Incorrect Answers

A global variable can mix different customers' items.

Unscoped persistent storage creates leakage and violates data minimization.

Pretraining does not contain a particular authenticated session's newly provided choice.

Question 161

An advisor needs preferences across separate sessions. Which change preserves actual tool authorization and trustworthy state?

Use the runtime's ephemeral microVM filesystem as the only cross-session preference database.
Copy every user's preferences into a single shared semantic-memory namespace.
Save indefinitely every raw conversation without consent or deletion controls.
Use consented long-term memory scoped to a verified identity and retention policy.

Explanation

Correct Answer

Use actor-scoped, consented long-term AgentCore Memory with retention limits; short-lived runtime sessions do not make durable preference stores.

Incorrect Answers

Runtime sessions terminate; ephemeral filesystem state is not durable long-term memory.

A common namespace can expose private preferences across actors.

Durable memory should be purpose-limited and governed by retention and user rights.

Question 162

A multi-tenant agent returns a different user's conversation memory. What should the developer implement before allowing the agent to act?

Accept an arbitrary actorId supplied by the prompt as permission to read those memories.
Search the entire shared memory store and ask the model not to quote other users.
Enforce actor/tenant isolation in AgentCore Memory before retrieval.
Remove the session ID and rely on similarity score to distinguish private memories.

Explanation

Correct Answer

Validate actorId/sessionId and namespace-level memory policies before lookup; a session key by itself does not prove ownership.

Incorrect Answers

Model/user text cannot authenticate a memory subject.

Retrieval has already exposed unauthorized data before any response filter.

Similarity cannot enforce actor identity isolation.

Question 163

A refund agent is orchestrated by a workflow that has generated a valid refund proposal. Financial policy requires a named manager to approve this particular refund before the payment service commits it. How should the orchestration enforce the pause and resume?

Pause before refund execution, obtain independently authenticated manager approval tied to this pending operation, then resume only after validating that approval at the payment commit boundary.
Treat the model's claim that a manager approved it as a signed approval.
Issue the refund and ask the manager for retroactive approval afterward.
Store an approval checkbox inside untrusted tool output and skip identity verification.

Explanation

Correct Answer

An agent’s plan or model text does not grant payment authority. Hold workflow execution before side effects; a separately authenticated approval bound to the pending refund must be checked before the service commits it.

Incorrect Answers

An FM has no independent approver credential.

The irreversible financial action would occur before the required approval.

Tool output cannot prove a manager actually authorized the refund.

Question 164

A tool reports malformed JSON arguments from a model. Which change preserves actual tool authorization and trustworthy state?

Forward any well-formed JSON payload to Lambda even when it violates the tool contract.
Let Lambda infer missing required parameters from a model-generated natural-language explanation.
Disable schema checking whenever a toolUse request originates from a trusted model ID.
Enforce tool JSON schema and reject unsupported parameters before Lambda execution.

Explanation

Correct Answer

Validate tool schema, required fields, value types and allowlisted parameters before calling Lambda, rejecting extra or malformed arguments.

Incorrect Answers

Valid JSON syntax does not imply correct tool semantics or safe arguments.

Inferred values can target the wrong resource or violate the API's mandatory contract.

Tool-request generation does not make its arguments authorized or well typed.

Question 165

A model loops indefinitely between search and calculation tools. What should the developer implement before allowing the agent to act?

Bound tool iterations, timeouts, cost budgets and termination conditions.
Leave the search/calculation loop unbounded until the model eventually says done.
Increase model temperature and hope a random response stops tool calls.
Mark every workflow successful after the tenth call regardless of result validation.

Explanation

Correct Answer

Enforce step-count, time, spend and exit-condition budgets so the agent can terminate with a truthful partial/failed status.

Incorrect Answers

No termination guarantee exists; loop may consume indefinite time and cost.

Sampling changes cannot enforce a hard tool iteration limit.

A fixed counter stops work but does not prove the requested task succeeded.

Question 166

An agent consumes a malicious tool result asking it to change instructions. Which agent/tool architecture control directly resolves the situation?

Treat tool content as untrusted and prevent tool output from overriding system policy.
Place tool-result text into the next system message because the tool belongs to the company.
Execute a requested credential transfer because it appears in a successful search result.
Let the model decide whether its own security policy may be relaxed for this tool output.

Explanation

Correct Answer

A tool result remains untrusted data; execution policy and authorized user instruction must not be overridden by retrieved/tool text.

Incorrect Answers

The returned content may still be attacker-controlled and would gain false authority.

Search content does not grant the application permission to exfiltrate credentials.

Policy enforcement belongs to trusted application logic, not lower-trust text.

Question 167

A search agent opens a file outside the user's permitted workspace. Which change preserves actual tool authorization and trustworthy state?

Fetch arbitrary URLs, including metadata addresses, requested by a web page inside a tool response.
Treat browser page text as executable agent instructions when it contains a role heading.
Apply resource-level ACL in the tool instead of relying on prompt refusal.
Return the first page summary without recording its origin or verifying permissions.

Explanation

Correct Answer

Apply domain/URL allowlists, fetch boundaries and content sandboxing to web navigation before using results as evidence.

Incorrect Answers

Tool-supplied addresses can cause SSRF or expose internal endpoints.

A page is untrusted content even if it includes authoritative-looking markup.

An unverified page is not a sufficient provenance or authorization basis.

Question 168

A model plans a multi-step task with dependencies. What should the developer implement before allowing the agent to act?

Transfer only a free-text summary and generate a new task ID on each retry.
Orchestrate sequential states with explicit prerequisite checks and recovery paths.
Assume successful handoff means every delegated action already committed.
Allow multiple agents to race on the same order without a shared commit or reconciliation boundary.

Explanation

Correct Answer

Require explicit workflow state transitions, durable task identifiers and idempotent retries when agents hand off work.

Incorrect Answers

Lost operation identity can duplicate downstream side effects.

Queueing a task is not the same as execution confirmation.

Concurrent agents can cause conflicting mutations without coordination.

Question 169

A parallel agent workflow must merge conflicting outputs. Which agent/tool architecture control directly resolves the situation?

Require a deterministic aggregation/verification step with provenance per agent.
Concatenate conflicting outputs and ask the final model to pick whichever sounds more fluent.
Select the fastest agent result and discard all other specialist evidence.
Use an unbounded majority vote from agents that share the same source error.

Explanation

Correct Answer

Merge agent outputs through a deterministic verifier that records which agent/source supplied each claim and resolves contradictions explicitly.

Incorrect Answers

Fluency is not an authoritative conflict-resolution rule.

Latency ranking does not establish factual correctness.

Correlated mistakes can win a vote; verify the actual evidence and business rule.

Question 170

A supervisor chooses between code, search and finance specialists. Which change preserves actual tool authorization and trustworthy state?

Allow the supervisor to grant finance privileges just by naming the finance specialist.
Send every task to every specialist, then run all suggested actions in parallel.
Route by bounded capability contracts and authorize each specialist separately.
Let a model select a privileged specialist based on untrusted retrieved role descriptions.

Explanation

Correct Answer

Route specialist calls by documented capability and require independent resource authorization at each specialist/tool boundary.

Incorrect Answers

Tool routing does not create employee financial permissions.

Unnecessary action paths expand costs and unauthorized side-effect risk.

Retrieved text is not an authoritative capabilities or identity registry.

Pass Your Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Exam Easy!

Question 160

Explanation

Correct Answer

Incorrect Answers

Question 161

Explanation

Correct Answer

Incorrect Answers

Question 162

Explanation

Correct Answer

Incorrect Answers

Question 163

Explanation

Correct Answer

Incorrect Answers

Question 164

Explanation

Correct Answer

Incorrect Answers

Question 165

Explanation

Correct Answer

Incorrect Answers

Question 166

Explanation

Correct Answer

Incorrect Answers

Question 167

Explanation

Correct Answer

Incorrect Answers

Question 168

Explanation

Correct Answer

Incorrect Answers

Question 169

Explanation

Correct Answer

Incorrect Answers

Question 170

Explanation

Correct Answer

Incorrect Answers

AWS Certified Generative AI Developer - Professional AIP-C01 Practice Tests