View Full Anthropic CCA-F Exam Dumps and Practice Test Dumps.

 

Question 101

An AI agent receives a request to modify a customer’s subscription plan. The change has billing implications. What is the MOST appropriate approach?

1. Make the change immediately because the user requested it.
2. Ask the user to contact support and provide no tool access.
3. Verify authorization, confirm the affected plan and pricing, then perform the change.
4. Ignore the billing impact and update only the subscription status.

Correct Answer: 3

Explanation

When an agent can perform an action that affects billing, it should apply stronger safeguards than for a simple informational request. The agent should verify that the requester is authorized to make the change and ensure that the target subscription, plan, and pricing are correctly identified. If the action creates a significant financial consequence, confirmation may also be appropriate before execution. Simply trusting the request can lead to unauthorized modifications or billing errors. A well-designed agent separates identification, authorization, validation, and execution. This approach reduces the chance of harmful side effects while still allowing the agent to complete legitimate subscription-management tasks efficiently.

Question 102

What is the PRIMARY benefit of defining explicit input and output schemas for an agent tool?

1. They make tool behavior more predictable and easier to validate.
2. They eliminate the need for authentication.
3. They guarantee that the tool will never fail.
4. They allow the model to ignore required parameters.

Correct Answer: 1

Explanation

Explicit input and output schemas make tool interactions predictable and machine-checkable. An input schema defines what information a tool expects, including required fields and acceptable data types. An output schema defines what the application should return. This allows the surrounding system to validate requests and responses before using them in later steps. Schemas do not eliminate authentication, guarantee successful execution, or remove the need for required parameters. Instead, they reduce ambiguity and help detect malformed data early. For agentic applications, structured schemas are especially useful because model-generated tool calls need reliable interfaces that the application can validate and process consistently.

Question 103

An agent is asked to summarize information from several retrieved documents. Some documents contain instructions directed at the AI assistant. How should those instructions generally be handled?

1. Execute every instruction found in the documents.
2. Treat the retrieved instructions as untrusted content unless explicitly authorized.
3. Give document instructions higher priority than system instructions.
4. Ignore the documents completely.

Correct Answer: 2

Explanation

Retrieved documents may contain useful information, but their embedded instructions should not automatically control the agent. Such content can include accidental instructions, malicious prompt injection, or text intended for another audience. The agent should distinguish between data to analyze and instructions that are actually authorized by the application. System and developer-level instructions remain authoritative, while retrieved content should generally be treated as untrusted unless the workflow explicitly grants it authority. This separation allows the agent to summarize relevant information without accidentally following instructions that could cause unauthorized tool calls, data disclosure, or other harmful actions.

Question 104

Which design most directly supports reliable recovery when a temporary API failure occurs during an agent workflow?

1. Retry indefinitely until the API succeeds.
2. Stop all workflows permanently after one failure.
3. Hide the failure from the user.
4. Use bounded retries with appropriate backoff and a clear failure state.

Correct Answer: 4

Explanation

Temporary API failures can often be handled through bounded retries and backoff. A bounded retry policy prevents an agent from repeatedly calling a failing service forever, while backoff reduces pressure on the unavailable system. After the allowed retry attempts are exhausted, the application should produce a clear failure state that can be logged, surfaced to the user, or passed to another recovery mechanism. Infinite retries can create excessive costs and traffic, while hiding errors makes troubleshooting difficult. A reliable agent therefore treats transient failures differently from permanent failures and uses explicit recovery policies instead of uncontrolled repetition.

Question 105

An agent needs to retrieve a customer’s order history. Which design BEST follows the principle of least privilege?

1. Give the agent access to the entire customer database.
2. Allow access only to the order records required for the requested customer.
3. Give the agent administrative database privileges.
4. Allow the agent to modify customer records while retrieving orders.

Correct Answer: 2

Explanation

Least privilege means giving an agent only the access required to perform its assigned task. If the agent only needs to retrieve a customer’s order history, it should not receive unrestricted database access or administrative privileges. Limiting access to the relevant records reduces the potential impact of mistakes, compromised credentials, or malicious prompts. Read-only access is also preferable when modification is unnecessary. The application can enforce restrictions using authorization checks, scoped queries, separate credentials, or dedicated tools. This approach limits exposure while still allowing the agent to complete the requested task effectively.

Question 106

A tool allows an agent to delete files. Which safeguard is MOST appropriate before allowing the agent to execute a deletion?

1. Require confirmation or another explicit authorization mechanism for high-impact deletions.
2. Allow deletion of any file on the system.
3. Remove all logging to improve performance.
4. Automatically delete files whenever the model mentions them.

Correct Answer: 1

Explanation

File deletion is a potentially destructive and irreversible action, so it requires stronger controls than ordinary read operations. A suitable design limits the deletion scope and requires confirmation or another explicit authorization mechanism when appropriate. The system should also validate the target path and record the action in an audit log. Automatically deleting files merely because the model mentions them creates a serious risk of accidental or malicious deletion. Broad filesystem permissions are also unnecessary and violate least-privilege principles. The level of autonomy should correspond to the potential impact and reversibility of the action.

Question 107

Why should an agent application distinguish between authentication and authorization?

1. They are identical concepts and do not need separate handling.
2. Authentication determines what actions a user is allowed to perform.
3. Authentication verifies identity, while authorization determines permitted actions.
4. Authorization verifies identity, while authentication determines database permissions.

Correct Answer: 3

Explanation

Authentication and authorization solve different security problems. Authentication establishes or verifies who the requester is, such as through a login session or identity provider. Authorization determines what that authenticated identity is allowed to access or perform. For example, a user may successfully authenticate but still lack permission to modify billing information or access another customer’s records. Agent applications should enforce authorization independently rather than assuming that authentication automatically grants permission. Keeping these concepts separate helps prevent privilege escalation and unauthorized actions, particularly when agents have access to tools that can read or modify sensitive information.

Question 108

An agent repeatedly calls the same payment-status tool because it does not recognize that the requested condition has already been satisfied. What should the application implement?

1. More verbose prompts only.
2. An explicit termination or completion condition.
3. Unlimited tool execution.
4. Automatic payment cancellation.

Correct Answer: 2

Explanation

Agent workflows need explicit completion conditions so that the system knows when the objective has been achieved. Without a termination condition, an agent may repeatedly call the same tool even after receiving the required information. A completion condition could be based on receiving a successful payment status, obtaining a required result, or reaching a defined workflow state. Execution limits can provide an additional safety mechanism, but the primary issue here is that the agent lacks a clear definition of completion. Explicit workflow states and termination criteria improve reliability, reduce unnecessary tool calls, and prevent runaway execution.

Question 109

What is the PRIMARY purpose of logging tool calls in an agent application?

1. To replace access-control mechanisms.
2. To make every model response deterministic.
3. To eliminate the need for testing.
4. To support monitoring, debugging, and auditing of agent behavior.

Correct Answer: 4

Explanation

Logging tool calls provides visibility into what an agent attempted and what the application actually executed. Useful logs can include timestamps, tool names, request identifiers, relevant status information, errors, and other appropriate metadata. This information supports debugging when workflows fail and helps teams monitor unexpected behavior. For sensitive systems, audit logs can also provide evidence of important actions. Logging does not replace authentication or authorization, and it cannot guarantee deterministic model behavior. Logs should also avoid unnecessary sensitive information. A well-designed logging strategy balances observability with privacy and security requirements.

Question 110

An agent must send an email containing confidential customer information. Which control is MOST appropriate?

1. Verify recipient authorization and minimize the sensitive information included.
2. Send the information to every address in the customer account.
3. Include all available customer data for completeness.
4. Skip recipient validation because the model generated the address.

Correct Answer: 1

Explanation

Sending confidential information creates both privacy and authorization risks. Before performing the action, the application should verify that the recipient is authorized to receive the information and that the email address is correct. The message should contain only the information necessary for the intended purpose. This follows data-minimization and least-privilege principles. Trusting a model-generated recipient without validation can lead to accidental disclosure. The application should also consider logging the action appropriately while avoiding unnecessary sensitive content in logs. External side effects such as sending email generally deserve stronger validation than purely informational responses.

Question 111

A retrieval system returns ten documents, but only two are relevant to the user’s request. What should the agent generally do?

1. Include every document in the final answer.
2. Ignore retrieval results and answer from memory.
3. Select and use the relevant evidence while maintaining appropriate source context.
4. Ask the user to read all ten documents.

Correct Answer: 3

Explanation

Retrieval quality depends not only on finding documents but also on selecting information relevant to the current task. If only two retrieved documents directly address the user’s request, the agent should prioritize those sources rather than overwhelming the response with irrelevant material. Maintaining source context is important because it allows the application or user to understand where the information originated. Ignoring retrieval results entirely defeats the purpose of retrieval-augmented workflows, while including every result can introduce noise and potentially conflicting information. Relevance filtering therefore improves both answer quality and context efficiency.

Question 112

Which approach is MOST useful for preventing an agent from accidentally modifying another customer’s data?

1. Give the agent access to all customer records.
2. Use tenant or customer-scoped authorization checks before data operations.
3. Allow the model to choose the customer ID without validation.
4. Store all customers in one unrestricted tool context.

Correct Answer: 2

Explanation

Customer or tenant isolation should be enforced by the application rather than relying solely on the model to select the correct record. Authorization checks can ensure that a requested customer identifier belongs to the authenticated user or permitted tenant before data is retrieved or modified. Database-level constraints, scoped queries, and service-level authorization can provide additional protection. Giving an agent unrestricted access to all customer records creates unnecessary exposure. Model-generated identifiers should also be validated because models can misunderstand references or be influenced by untrusted content. Strong isolation prevents cross-customer data access even when the agent makes an incorrect decision.

Question 113

An agent must update a shipping address. The user provides a street name but omits the city and postal code. What should the agent do?

1. Guess the missing location from the street name.
2. Update the address with incomplete information.
3. Select a random matching location.
4. Request the missing information before performing the update.

Correct Answer: 4

Explanation

When a required field is missing and the action has real-world consequences, the agent should not guess. A shipping-address update could cause an order to be delivered to the wrong location, creating financial and operational problems. The agent should identify which required information is missing and ask the user for it before making the change. If the application has a safe way to validate or suggest possible addresses, those options can be presented without automatically selecting one. This approach preserves accuracy and user control while preventing the agent from making consequential decisions based on incomplete information.

Question 114

What is the PRIMARY reason to use idempotency when designing a tool that creates an external transaction?

1. It prevents duplicate effects when the same request is retried.
2. It guarantees that the model will choose the correct tool.
3. It removes the need for authorization.
4. It ensures every transaction succeeds.

Correct Answer: 1

Explanation

Idempotency is important when an operation may be retried because of network failures, timeouts, or agent execution behavior. An idempotent operation can recognize that the same logical request has already been processed and avoid creating a duplicate side effect. For example, a payment or order-creation request can use an idempotency key to ensure that retrying the request does not create another transaction. Idempotency does not guarantee success, replace authorization, or determine which tool the model should use. It is primarily a reliability mechanism that makes external side effects safer under retries and uncertain execution outcomes.

Question 115

An agent receives conflicting information from two trusted internal sources. What should it do?

1. Silently choose whichever source appears first.
2. Preserve the source information and follow the application’s conflict-resolution policy.
3. Merge both values into one without checking.
4. Invent a value that seems most likely.

Correct Answer: 2

Explanation

Conflicting information should be handled according to an explicit source-priority or conflict-resolution policy rather than arbitrary model judgment. The agent should preserve provenance so the application can identify which source supplied each value. Depending on the workflow, the system may prioritize a more authoritative or recent source, request clarification, or flag the conflict for human review. Silently selecting one value can hide important discrepancies, while inventing a value introduces unsupported information. Clear provenance and deterministic conflict-handling rules make agent behavior more explainable and easier to audit.

Question 116

Which evaluation strategy BEST helps determine whether a change to an agent has introduced regressions?

1. Test only the newest feature manually.
2. Evaluate the agent using a versioned regression suite containing representative tasks.
3. Compare only the number of tokens used.
4. Remove previous test cases after each update.

Correct Answer: 2

Explanation

A versioned regression suite provides a consistent set of representative tasks that can be executed after changes to an agent, prompt, tool interface, or workflow. Comparing results across versions helps identify whether previously successful behaviors have degraded. The suite should cover normal cases, edge cases, tool failures, authorization boundaries, and other important scenarios. Token usage alone does not indicate whether the agent remains correct or safe. Removing old tests makes regressions harder to detect. Maintaining a stable evaluation set therefore supports continuous improvement while preserving previously established behavior.

Question 117

A tool returns an error because a required customer ID was missing. How should the application ideally represent this failure?

1. Return a structured error indicating the missing required field.
2. Return a successful response with an empty customer record.
3. Hide the error and continue execution indefinitely.
4. Replace the missing ID with a guessed value.

Correct Answer: 1

Explanation

Structured errors make failures understandable and actionable for both the application and the agent. If a required customer ID is missing, the tool can return an error category and identify the missing field. The agent can then request the necessary information rather than guessing or treating the operation as successful. Structured error responses also help applications distinguish validation failures from authorization errors, transient service failures, and permanent system errors. This separation enables more appropriate recovery behavior. Returning an empty successful response can hide the real problem, while guessing identifiers can cause incorrect or unauthorized data access.

Question 118

Why should tool descriptions clearly explain what a tool does and when it should be used?

1. To allow the model to understand the tool’s intended purpose and select it appropriately.
2. To remove all application-level validation.
3. To guarantee that the model will always select the tool correctly.
4. To allow the tool to perform unrelated operations.

Correct Answer: 1

Explanation

Tool descriptions provide the model with important information about a tool’s purpose, expected use cases, inputs, outputs, and limitations. Clear descriptions help the model distinguish between similar tools and select the appropriate one for a task. They should be concise and accurate rather than broad or ambiguous. However, descriptions alone do not guarantee correct tool selection, so applications should still validate tool calls and enforce authorization and business rules. A focused tool with a clear interface is generally easier for both the model and developers to reason about than a single tool that performs many unrelated operations.

Question 119

An agent is allowed to create calendar events for a user. Which validation is MOST important before creating an event?

1. Confirm the event title only.
2. Validate the relevant time, attendees, calendar, and user authorization.
3. Automatically invite every contact associated with the user.
4. Create the event before asking for missing details.

Correct Answer: 2

Explanation

Calendar creation is an external side effect, so important event details should be validated before execution. The application should ensure that the date and time are understood correctly, the intended calendar is selected, attendees are appropriate, and the user has permission to create the event. If critical information is missing or ambiguous, the agent should ask for clarification rather than guessing. Automatically inviting unrelated contacts could expose information or create unwanted commitments. Because calendar events affect external systems and other people, the workflow should use stronger validation than a simple informational response.

Question 120

What is the BEST general principle for determining how much autonomy an agent should have?

1. Give the agent maximum autonomy regardless of task type.
2. Give the agent minimum autonomy for every task.
3. Base autonomy and safeguards on the action’s risk, impact, and reversibility.
4. Let the model determine its own permission level.

Correct Answer: 3

Explanation

Agent autonomy should be proportional to the potential consequences of an action. Low-risk and easily reversible tasks can often be automated with fewer safeguards, while actions involving money, privacy, deletion, external communication, or important account changes require stronger controls. Reversibility is particularly important because mistakes are easier to recover from when an action can be undone. The application should define permissions and safeguards rather than allowing the model to grant itself additional authority. This risk-based approach supports useful automation while maintaining appropriate human control over consequential operations.