Question 80
A user wants an agent to handle weekly unrelated support cases without leaking old customer data. What is the most suitable session policy?
- Rely on a final instruction saying previous cases are irrelevant
- Fork every old customer session into a shared weekly master
- Create a fresh case-specific session and load only scoped case records
- Reuse the previous customer's live session for all cases
Explanation
Correct Answer
C:New sessions isolate customers and reduce stale context. Long-lived knowledge should reside in authorized systems and be fetched for each case.
Incorrect Answers
A:Prompts cannot guarantee absence of sensitive information already present in the context. Explicit isolation is preferable.
B:Forks would carry historical content into a common context. This defeats case-level separation.
D:That design risks exposing private data and mixing case facts. Session continuity is inappropriate across unrelated customers.
Question 81
A paused synthesis agent is resumed after one of its sources has published a corrected figure. What should determine whether to reuse its prior conclusion?
- Publish the old conclusion unchanged because the session is resumable
- Validate source freshness and recompute any claim depending on the corrected data
- Discard every other source regardless of whether it changed
- Assume a fresh session would automatically discover the correction without retrieval
Explanation
Correct Answer
B:Session history captures an earlier reasoning state, not an immutable truth. Recovery should identify facts affected by new evidence.
Incorrect Answers
A:Resumability does not guarantee external evidence is still valid. Stale figures can make the answer wrong.
C:A targeted revision is preferable to deleting unrelated valid evidence. Only dependent claims need reevaluation.
D:Starting fresh does not guarantee current-source verification. Explicit evidence checking is still required.
Question 82
A travel agent has tools named search_flights and search_hotels, but it often calls the flight tool when asked for lodging. What change most directly improves routing?
- Increase the maximum output tokens for every search
- Give both tools unrestricted access to every booking database
- Differentiate tools by intent, parameters and boundaries
- Rename both tools to search without changing their descriptions
Explanation
Correct Answer
C:Claude uses tool documentation to decide which capability fits a request. Descriptions should identify lodging versus flights and specify when not to choose each.
Incorrect Answers
A:Generation length is not the central issue. The model needs to distinguish the correct external capability before it executes.
B:Expanded privileges do not make the functions more distinguishable. They increase risk and mask a semantic interface problem.
D:Ambiguous identical names make selection harder. A shared generic label removes a useful discriminator rather than adding one.
Question 83
A CRM lookup tool takes an account_id string but callers frequently provide customer names. Which schema improvement addresses the input contract?
- Allow arbitrary JSON fields and silently ignore unknown inputs
- Specify the required identifier format in input_schema
- Remove the parameter description to save a few tokens
- Make every field optional and guess the account when missing
Explanation
Correct Answer
B:Precise parameter definitions help the model supply the proper key. Runtime validation should still reject names where a canonical ID is required.
Incorrect Answers
A:Loose parsing can conceal malformed calls and increase failures. A narrow schema provides clearer feedback and safety.
C:Less guidance makes the correct argument harder to infer. The problem is an insufficiently specified interface.
D:A permissive schema can hide invalid or ambiguous requests. Account selection should be grounded in verified identifiers.
Question 84
A service offers get_invoice and refund_invoice tools. The model chooses refund_invoice when asked only to view an invoice. What design change is most useful?
- Add more unused parameters to the refund tool
- Ask users to approve every read-only invoice lookup manually
- Merge both methods into a single unspecified invoice tool
- Document read versus write effects in both tool descriptions
Explanation
Correct Answer
D:The interface should distinguish retrieval from financial actions. Clear intent boundaries reduce the chance of calling a destructive operation for a read-only request.
Incorrect Answers
A:Unrelated arguments do not teach the model when to use a capability. They can make the contract harder to satisfy.
B:Extra approval does not correct misselection of write versus read. The tools need better naming and boundaries.
C:An overloaded operation blurs the difference further. Sensitive writes should have explicit semantics and authorization.
Question 85
A repository tool exposes 50 tiny operations and the agent repeatedly selects the wrong file-inspection method. What consolidation is most appropriate?
- Replace repository tools with direct mutation permission only
- Expose all 50 with only numeric names to shorten documentation
- Group related operations with clear inputs and bounded outputs
- Force the agent to guess a function by position in the list
Explanation
Correct Answer
C:A focused catalog makes choice easier and reduces context overhead. Operations should map to meaningful jobs rather than every internal implementation step.
Incorrect Answers
A:Editing access cannot substitute for reliable inspection. The user needs safe retrieval first.
B:Opaque names remove semantic cues and make the selection problem worse. Token savings alone do not justify unusable interfaces.
D:Tool order is not a dependable semantic contract. The request should map to an understandable capability description.
Question 86
A search tool returns raw debug logs, internal tokens and one relevant answer. How should its output be redesigned for an agent consumer?
- Return relevant findings, stable references and status
- Return all debug traces so the model has maximal context
- Make the output a random natural-language story
- Return only the word done without evidence
Explanation
Correct Answer
A:Agent-facing output should surface actionable facts. Removing secrets and irrelevant logs reduces context burden while keeping traceability.
Incorrect Answers
B:More text can bury the answer and expose secrets. Context should be curated for the consuming task.
C:Narrative embellishment introduces uncertainty and parsing difficulty. A predictable result contract is more useful.
D:A success marker without results cannot support a grounded answer. The consumer needs the actual finding and source.
Question 87
A purchase-order tool accepts currency but does not state supported values, causing inconsistent calls. What is the best interface fix?
- Remove currency from the tool and let the agent decide later
- Add a lengthy essay about corporate finance unrelated to the field
- Accept any string and convert unknown codes into USD silently
- Constrain supported currencies with schema enums or clear validation rules
Explanation
Correct Answer
D:An explicit set of values improves argument generation and validation. Unsupported currency codes can then receive actionable errors.
Incorrect Answers
A:Currency is part of the order semantics. Omitting it can make amounts ambiguous or incorrect.
B:General prose does not specify the contract. The schema needs concrete allowed values and behavior.
C:Silent coercion can change financial meaning. Callers need to know when a requested currency is unsupported.
Question 88
A tool catalog contains similar invoice-reading and refund-changing capabilities. Which TWO interface details most help Claude choose safely between them? Choose TWO.
- Opaque names such as action1 and action2 without descriptions
- Precise required parameters and validation boundaries for each operation
- Clear descriptions that identify read-only versus mutation behavior
- Examples that depict successful refunds for every possible user request
- Shared unrestricted financial permissions for every tool user
Explanation
Correct Answers
B:An explicit schema defines what identifiers and values a call needs. It enables the host to reject ambiguous or malformed requests.
C:Capability documentation should explain side effects and suitable intent. This helps avoid initiating a refund when the user only asks for an invoice.
Incorrect Answers
A:Removing semantic hints makes function selection harder. The model cannot reliably infer risk and purpose from numbers.
D:Misleading demonstrations blur the distinction between read and write tasks. They can encourage improper mutation.
E:Broad access is not tool documentation and undermines least privilege. It does not improve routing accuracy.
Question 89
An agent often searches a full database to retrieve one invoice by exact identifier. What tool capability would better match this common task?
- Disable the identifier field so all requests become free text
- Force a model to summarize the whole database first
- Provide an exact-lookup operation with a required invoice identifier
- Increase search result count on the generic database tool
Explanation
Correct Answer
C:Task-aligned tools reduce unnecessary broad retrieval and produce more predictable results. The interface can enforce identifier validation.
Incorrect Answers
A:Removing a useful constraint weakens precision. A known invoice ID should be used directly.
B:Global summarization is expensive and unnecessary for exact-key retrieval. It also increases exposure of unrelated records.
D:More matches add noise and make correct selection harder. The request already has a precise key.
Question 90
A legal-research tool is advertised as providing definitive legal judgments, but it actually retrieves web snippets. What should its description say?
- Rename the tool to final_verdict without changing implementation
- State that retrieval does not certify legal conclusions
- Keep the authoritative claim because it increases tool usage
- Hide source locations to make the output appear cleaner
Explanation
Correct Answer
B:Tool descriptions are the contract for model planning. A retrieval tool should represent evidence access accurately and preserve uncertainty.
Incorrect Answers
A:Naming does not confer expert validation. The mismatch between capability and description would persist.
C:Overclaiming capabilities can cause unsupported legal assertions. The agent may confuse retrieved text with verified advice.
D:Removing citations makes claims harder to verify. Legal research especially needs provenance.