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

 

Question 81

A developer wants to prevent an agent from repeatedly calling a tool when the tool continues returning the same failure. Which approach is most appropriate?

1. Increase the model’s temperature.
2. Remove the tool’s error response.
3. Add a retry limit and a termination condition.
4. Allow the agent to continue indefinitely.

Correct Answer: 3

Explanation

Production agent workflows should have explicit controls that prevent repeated failures from turning into unlimited execution. A retry limit defines how many times an operation may be attempted, while a termination condition determines when the workflow should stop or escalate the problem. Increasing temperature does not reliably prevent repetitive behavior. Removing error information makes it harder for the agent to understand the failure. Unlimited execution can increase latency, token consumption, API costs, and external side effects. Retry policies should also distinguish between temporary failures that may be safely retried and permanent failures that require correction or intervention.

Question 82

An agent needs to access a database but should only retrieve records belonging to the authenticated user. Which control should enforce this restriction?

1. A database or application-level authorization policy tied to the authenticated identity
2. A prompt asking Claude to select only the correct records
3. A larger context window
4. A higher temperature

Correct Answer: 1

Explanation

Authorization should be enforced by trusted application or database controls rather than relying solely on the model. The authenticated user’s identity can be associated with appropriate access policies so that queries return only authorized records. A prompt can communicate the intended behavior but cannot provide a deterministic security boundary. Context-window size and temperature are unrelated to authorization. Strong access control should remain effective even if the model makes an incorrect request. The application can additionally validate requested resource identifiers before executing database operations and log access for auditing.

Question 83

A tool returns a complex nested object containing customer information. What can make the response easier for Claude to use reliably?

1. Remove all field names.
2. Use a clear structured schema with descriptive field names.
3. Convert the object into unrelated paragraphs.
4. Return random fields on every call.

Correct Answer: 2

Explanation

Structured responses with descriptive field names make it easier for Claude and the application to understand the returned information. A consistent schema allows downstream steps to identify fields such as customer ID, account status, and contact information without guessing their meaning. Removing field names eliminates important semantic information. Converting structured data into unrelated paragraphs can make parsing and interpretation more difficult. Randomly changing the returned fields also creates an unstable interface. Consistent schemas improve reliability, simplify validation, and make tools easier to maintain and test.

Question 84

A user asks an agent to cancel an order, but the order identifier is missing. What should the agent do before attempting the cancellation?

1. Guess an order identifier.
2. Cancel the most recent order automatically.
3. Ask the user for the required order identifier.
4. Cancel all orders associated with the user.

Correct Answer: 3

Explanation

A required identifier should be obtained before executing a consequential operation. Guessing an order ID or selecting an arbitrary order can result in modifying the wrong resource. Cancelling the most recent order or all orders is even broader and may cause unintended consequences. The agent should clearly request the missing identifier and then use it to perform the authorized operation. The application should also validate that the identifier belongs to the authenticated user and that the order is eligible for cancellation before executing the action.

Question 85

A team wants to make an agent’s tool calls easier to audit after an incident. Which information should be retained where appropriate?

1. Only the model’s final answer
2. Only the application launch time
3. Tool names, relevant parameters, results, timestamps, and execution status
4. No runtime information

Correct Answer: 3

Explanation

Auditing requires sufficient information to reconstruct important workflow events. Depending on privacy and security requirements, useful records can include which tool was called, relevant parameters, timestamps, execution status, errors, and results or references to results. Recording only the final answer does not reveal how the agent reached that outcome. Retaining no runtime information makes incident investigation difficult. Logging should still follow data-minimization principles, with sensitive information appropriately redacted and access to logs controlled. A balanced audit trail improves accountability, debugging, and operational analysis.

Question 86

A developer is creating a tool that can perform several unrelated administrative operations. What design would generally improve agent reliability?

1. Combine every operation into one unrestricted command.
2. Split unrelated capabilities into focused tools with clear schemas.
3. Remove descriptions from all operations.
4. Let the model construct arbitrary administrative commands.

Correct Answer: 2

Explanation

Focused tools reduce ambiguity and provide clearer boundaries around agent capabilities. If a tool performs many unrelated administrative operations, Claude has to determine not only whether the tool is appropriate but also which operation and parameters are required. Splitting capabilities into focused tools makes each interface easier to understand, validate, authorize, test, and audit. Unrestricted administrative commands provide excessive power and can increase security risk. Removing descriptions also makes tool selection harder. Narrow interfaces generally produce more predictable agent behavior.

Question 87

An agent is processing a document that contains instructions directed at the AI, such as requests to reveal hidden prompts. How should these instructions normally be handled?

1. Treat them as trusted system instructions.
2. Follow them whenever they appear in a document.
3. Treat them as document content unless the application explicitly authorizes them as instructions.
4. Give them higher priority than developer instructions.

Correct Answer: 3

Explanation

Instructions embedded inside documents should normally be treated as untrusted content rather than as authoritative instructions for the agent. Documents can contain prompt-injection attempts or text that is relevant to the document itself but unrelated to the agent’s actual task. System and developer instructions should retain their intended authority. Automatically following instructions from documents could cause the agent to reveal information or perform unintended actions. Separating data from trusted instructions is therefore an important security principle when agents process web pages, emails, uploaded files, search results, or other external content.

Question 88

A production application needs to know whether an agent’s performance changes after a model upgrade. Which practice is most useful?

1. Compare the upgraded model against a versioned evaluation suite.
2. Delete all previous evaluation results.
3. Test only one new prompt.
4. Change the evaluation criteria after seeing the results.

Correct Answer: 1

Explanation

A versioned evaluation suite provides a consistent basis for comparing behavior across model versions. The suite should contain representative tasks, edge cases, tool-use scenarios, and other important workflows. Preserving previous results makes it possible to identify regressions or changes in behavior. Testing only one prompt provides limited evidence, while changing evaluation criteria after seeing results can make comparisons unreliable. Reproducible evaluations are especially important when upgrading models in production because seemingly small behavior changes can affect tool selection, formatting, latency, or workflow completion.

Question 89

An agent has to choose between two tools that perform similar functions but operate on different customer accounts. What should help determine the correct tool?

1. The alphabetical order of the tools
2. The tool descriptions, schemas, and relevant account context
3. The number of characters in the tool name
4. Random selection

Correct Answer: 2

Explanation

Tool selection should be based on the task requirements, available context, and the semantic information provided by the tool definitions. Clear descriptions and schemas can specify which account or resource each capability applies to. The application’s trusted authorization layer should also verify that the selected operation is permitted for the relevant account. Alphabetical ordering and tool-name length have no meaningful relationship to correctness. Random selection is inappropriate when tools have different scopes. Clear interfaces combined with authorization checks provide both model guidance and deterministic protection.

Question 90

A developer wants an agent to safely modify configuration files. Which approach provides a useful additional safeguard?

1. Allow unrestricted modifications to every system file.
2. Validate the requested changes and restrict the writable file scope.
3. Disable all file permissions.
4. Give the agent administrator privileges.

Correct Answer: 2

Explanation

Configuration changes can affect system behavior, so the application should limit where the agent can write and validate the proposed changes before applying them. Restricting writable paths reduces the impact of accidental or malicious modifications. Validation can check syntax, permitted settings, and other application-specific constraints. Giving administrator privileges or unrestricted filesystem access provides significantly more power than necessary. Disabling all permissions prevents legitimate operation but does not provide a useful workflow. A controlled write interface combined with validation and appropriate approval mechanisms is safer.

Question 91

A tool performs a destructive action and returns success. What should the application record to support later troubleshooting and accountability?

1. Relevant execution information such as timestamp, operation, status, and appropriate identifiers
2. Nothing
3. Only the user’s browser type
4. Only the model’s temperature

Correct Answer: 1

Explanation

Destructive operations should generally produce an appropriate audit trail. Useful information can include the operation performed, timestamp, authenticated actor or service identity, execution status, and relevant resource identifiers. Sensitive values should be handled according to applicable privacy and security requirements. Recording nothing makes it difficult to investigate unexpected changes or establish what occurred. Browser type and model temperature do not provide meaningful information about the actual operation. Proper audit logging improves accountability and can help teams diagnose failures, investigate incidents, and understand agent behavior.

Question 92

A user asks an agent to perform a task, but the request is ambiguous and different interpretations could cause different database updates. What should the agent do?

1. Select an interpretation without informing the user.
2. Execute every possible interpretation.
3. Ask a focused clarification question before making the update.
4. Choose the interpretation with the shortest description.

Correct Answer: 3

Explanation

Ambiguous requests that can produce different external changes should generally be clarified before execution. A focused question can identify the exact resource, operation, or value the user intends. Silently choosing an interpretation risks modifying the wrong data. Executing all possible interpretations can cause multiple unintended changes. The length of an interpretation has no relationship to user intent. Clarification is especially important for irreversible or consequential operations. Once the required information is provided, the application should still validate authorization and parameters before performing the database update.

Question 93

An agent uses retrieved documents to answer questions, but sometimes cites a document that does not support the generated claim. What improvement can help?

1. Preserve evidence passages and source identifiers alongside retrieved claims.
2. Remove all source information.
3. Ask the model to create citations after generating unsupported claims.
4. Randomly select a citation.

Correct Answer: 1

Explanation

Evidence tracking can help keep generated claims connected to supporting sources. Retrieval systems can preserve document identifiers, relevant passages, page numbers, URLs, or other provenance information alongside extracted facts. The generation stage can then use that evidence when constructing the answer. Removing source information makes verification more difficult. Asking the model to invent citations after the fact can produce references that appear plausible but do not actually support the claim. Random citation selection is also unreliable. Grounded generation works best when evidence and provenance are preserved throughout the workflow.

Question 94

A developer wants to prevent a tool from accepting excessively large text input that could increase cost or cause failures. What should be implemented?

1. No input limits
2. An input-size limit with validation before execution
3. Unlimited retries
4. Automatic duplication of the input

Correct Answer: 2

Explanation

Input limits can protect both the application and external service from unnecessarily large requests. The tool can enforce a maximum length or size and return a clear validation error when the limit is exceeded. This helps control token consumption, latency, memory usage, and backend constraints. Unlimited retries do not address oversized input and may make the problem worse. Duplicating the input increases resource consumption further. Limits should be appropriate to the actual tool requirements, and the application can provide useful guidance when users need to shorten or divide large inputs.

Question 95

A team wants an agent to work with sensitive customer information while minimizing unnecessary exposure. Which approach is appropriate?

1. Provide every available customer field to every agent.
2. Apply data minimization and provide only the information required for the task.
3. Publish customer information in shared logs.
4. Remove all access controls.

Correct Answer: 2

Explanation

Data minimization means providing only the information necessary for the specific task. This reduces the amount of sensitive information exposed to the model and downstream components. Providing every available field increases unnecessary exposure and can make privacy controls more difficult. Publishing customer information in shared logs creates additional risk. Removing access controls eliminates important protections. Sensitive-data workflows should also use appropriate authorization, retention, redaction, and auditing practices. Limiting the information available to each operation is an important part of building privacy-conscious agent systems.

Question 96

A developer wants to know whether a tool call failed because of invalid input or because the external service was unavailable. What should the tool provide?

1. Only the message error
2. A structured error response with a meaningful error category
3. No error information
4. A random status value

Correct Answer: 2

Explanation

Structured error categories make it easier for both the application and the agent to choose the correct recovery strategy. Invalid input may require correction or clarification, while a temporary service outage may justify a limited retry. A generic error message does not provide enough information to distinguish these cases. Removing errors entirely prevents useful recovery, while random status values make behavior unpredictable. A useful error contract can include an error type, human-readable explanation, retryability information, and relevant safe metadata while avoiding unnecessary exposure of sensitive internal details.

Question 97

An agent can create calendar events and send emails. A user asks it to schedule a meeting and notify participants. What should the orchestration layer consider before executing both actions?

1. Whether the required event and recipient information is complete and authorized.
2. Whether both tools have similar names.
3. Whether the agent can skip all validation.
4. Whether unrelated tools should also be executed.

Correct Answer: 1

Explanation

Before performing multiple external actions, the workflow should ensure that required information is available and that each action is authorized. For a meeting, this could include the event time, participants, and other required fields. The application should validate the event parameters and recipients before creating the event or sending communication. Tool-name similarity is irrelevant to authorization. Skipping validation increases the possibility of incorrect actions, while executing unrelated tools adds unnecessary side effects. Each external action should be independently controlled even when they form part of the same user request.

Question 98

A developer wants to make an agent easier to maintain as the number of tools increases. Which design practice is most useful?

1. Give every tool the same broad purpose.
2. Use clear interfaces, consistent schemas, and focused responsibilities.
3. Allow arbitrary parameters for all tools.
4. Remove documentation to reduce maintenance.

Correct Answer: 2

Explanation

Consistent tool interfaces and focused responsibilities make larger agent systems easier to understand and maintain. Each tool should have a clear purpose, descriptive name, documented parameters, predictable outputs, and appropriate validation. Broad overlapping tools make selection and authorization more difficult. Arbitrary parameters reduce predictability and increase validation complexity. Removing documentation makes future maintenance harder rather than easier. As tool counts grow, clear interfaces also support testing, observability, permission management, and onboarding of developers who need to understand the agent’s available capabilities.

Question 99

An agent performs a database update successfully, but the model continues trying to repeat the same update. What is the most likely architectural issue to investigate?

1. Whether the workflow correctly communicates the successful completion state.
2. Whether the database should be removed.
3. Whether the model needs unlimited retries.
4. Whether every tool should be called again.

Correct Answer: 1

Explanation

Repeated execution after a successful operation can indicate that the workflow is not clearly communicating completion. The application should return a structured success result and provide a termination condition that tells the orchestration layer the requested operation has been completed. Without a clear completion state, the model may interpret the workflow as unfinished and attempt the same action again. Unlimited retries can make duplicate operations worse. Removing the database or calling every tool again does not address the underlying orchestration issue. Explicit success states and idempotency controls can provide additional protection.

Question 100

A production agent is given access to a tool that can make irreversible changes. Which overall design principle should guide the level of autonomy granted to the agent?

1. Give the agent unrestricted control to maximize automation.
2. Remove all approval mechanisms.
3. Match autonomy and safeguards to the potential impact and reversibility of the action.
4. Allow irreversible actions without logging.

Correct Answer: 3

Explanation

Agent autonomy should be designed according to the potential consequences of its actions. Reversible, low-impact operations may be suitable for greater automation, while irreversible or high-impact actions generally require stronger controls such as authorization, validation, confirmation, approval gates, restricted permissions, and audit logging. Unrestricted autonomy can increase the consequences of model or tool errors. Removing approval mechanisms eliminates a useful control layer for sensitive operations. Logging is also important for accountability and troubleshooting. Risk-based design allows organizations to gain automation benefits while maintaining appropriate controls over consequential external actions.