Question 90
An agent proposes an operation name not registered by the application. What should the dispatcher do?
- Install a package named after the proposed operation.
- Evaluate the name as arbitrary application code.
- Choose the nearest spelling automatically for every operation.
- Reject unknown operations and return a controlled error.
Explanation
Correct Answer
D:The dispatcher should execute only explicitly registered capabilities. A generated name cannot create a new authorized function.
Incorrect Answers
A:A generated identifier does not authorize software installation or establish implementation trust.
B:This turns model output into unrestricted execution rather than constrained tool use.
C:A similar name may have different side effects; silent substitution can execute the wrong action.
Question 91
A user requests a dry run of a deployment plan. Which behavior preserves that mode?
- Run the deployment and roll it back if anything changes.
- Use supported preview or validation operations and report the proposed effects.
- Call the production commit endpoint with a note saying dry run.
- Permit writes that the model considers harmless.
Explanation
Correct Answer
B:A dry run should inspect or simulate without committing changes. Any preview limitations should remain explicit.
Incorrect Answers
A:Execution followed by rollback is not a dry run and may leave irreversible side effects.
C:A prose note does not change endpoint semantics unless a supported nonmutating mode exists.
D:The requested mode excludes actual deployment changes, not only changes the model predicts are harmful.
Question 92
An auditor must connect a final recommendation to the tools and observations that produced it. Which trace design helps most?
- Record correlated task, tool-invocation, result, and decision identifiers.
- Capture detailed tool payloads but omit the task and invocation identifiers that connect them.
- Store only the final answer text.
- Assign an ID to each event but provide neither parent links nor a shared trace identifier.
Explanation
Correct Answer
A:Trace linkage allows reviewers to follow evidence through the workflow. Sensitive payloads still need appropriate minimization and access controls.
Incorrect Answers
B:Payload detail alone can leave concurrent or retried operations indistinguishable. The audit needs explicit links from observations to the decisions that used them.
C:The answer alone omits the execution path and supporting observations.
D:Unique event names prevent name collisions but do not express relationships. The auditor still cannot reliably reconstruct which result informed a particular decision.
Question 93
A team wants to compare two orchestration policies using the same recorded incidents. What makes the comparison useful?
- Use matched incidents, fixed success criteria, and comparable tool conditions.
- Compare only how persuasive their final summaries sound.
- Allow the new policy unlimited resources while tightly limiting the baseline.
- Give the new policy easier incidents.
- Run each policy against whichever incidents happen to arrive that day and attribute every score difference to the orchestration change.
Explanation
Correct Answer
A:Controlled replay isolates architectural differences. Tool snapshots or simulations should preserve the conditions needed for a fair comparison.
Incorrect Answers
B:Presentation does not measure task success, safety, or execution cost.
C:Different budgets can explain the outcome independently of orchestration quality.
D:Different difficulty confounds the effect of the policy change.
E:Different input mix and tool conditions can explain the difference. A matched comparison controls those confounders before assigning causality to the policy.
Question 94
An agent's completion rate rises after it starts declaring success on unresolved cases. Which metric change exposes this behavior?
- Count every final answer as a successful completion.
- Stop recording unresolved cases.
- Require externally verifiable task outcomes in the success definition.
- Increase the weight of positive wording.
Explanation
Correct Answer
C:Self-reported completion can be gamed by the system's reporting behavior. Outcome-based checks distinguish accomplished work from optimistic labels.
Incorrect Answers
A:This repeats the flawed definition and rewards unsupported closure.
B:Excluding failures improves the reported number while hiding actual performance.
D:Sentiment does not establish whether the requested state was reached.
Question 95
A more complex agent improves accuracy slightly but triples latency beyond the product's service target. What should the architect evaluate?
- Only the cheapest model's advertised price.
- Accuracy alone, because latency is unrelated to architecture.
- The number of agents as the primary quality measure.
- Whether the measured quality gain justifies the latency and cost under the product constraints.
Explanation
Correct Answer
D:Architecture selection is a tradeoff against explicit requirements. A simpler or selectively escalated design may achieve a better operating point.
Incorrect Answers
A:End-to-end cost includes calls, retries, tools, and time, not merely per-token pricing.
B:Orchestration affects both resource use and responsiveness, which are product requirements.
C:More components do not inherently produce a better user outcome.
Question 96
Most requests are straightforward, while a small fraction need deeper reasoning. Which model-routing design can reduce average cost without ignoring difficult cases?
- Send every request to the most expensive model without measuring whether the extra cost improves simple cases.
- Send every request to the cheapest model and escalate only when its response contains no text, without measuring quality failures.
- Route solely by input length, treating every short request as easy without validating that rule.
- Escalate only when the baseline reports uncertainty, without evaluating cases where it is confidently wrong.
- Use a capable baseline with measured escalation criteria for difficult or failed cases.
Explanation
Correct Answer
E:Selective escalation allocates expensive reasoning where it adds value. The criteria and fallback behavior must be evaluated on representative traffic.
Incorrect Answers
A:Uniform escalation may meet quality needs but discards the stated cost opportunity. The design should establish where a less costly baseline meets the required standard.
B:A model can return fluent but inadequate text. Escalation criteria need evidence of difficult or failed task performance, not merely output presence.
C:Short inputs can contain difficult reasoning or high-consequence decisions. A useful escalation rule needs evidence that its signals detect the cases the baseline handles poorly.
D:Self-reported certainty can miss quality failures. Escalation criteria need measured performance on difficult cases rather than relying solely on the baseline to recognize its own mistakes.
Question 97
A controller executes a search tool but starts the next reasoning turn before the result arrives. What is the main risk?
- The search tool's parameters become invalid merely because another model turn starts.
- The next decision may be made without the observation it depends on.
- The completed search result will retroactively change decisions already made in the next turn.
- The controller can substitute a predicted result whenever the external search is slow.
Explanation
Correct Answer
B:An agent loop should incorporate required tool results before choosing dependent actions. Otherwise it can reason from stale or missing evidence.
Incorrect Answers
A:The described problem is a missing observation dependency, not an automatic change to argument validity.
C:A later observation does not automatically revise actions already chosen without it. The controller must respect the dependency before making that decision.
D:A predicted result is not the required observation. Using it would remove the evidentiary dependency instead of satisfying it.
Question 98
A primary lookup service fails. A proposed fallback returns older approximate data. What must the orchestrator consider before substitution?
- Whether the fallback has lower latency than the primary endpoint.
- Only whether the fallback returns the same JSON field names.
- Whether the fallback's freshness and semantics satisfy the original requirement.
- Substitute the fallback whenever it has the same provider, without checking its data contract.
Explanation
Correct Answer
C:A reachable service is not necessarily an equivalent source. The system must disclose or reject degraded evidence when it cannot meet the task's constraints.
Incorrect Answers
A:Faster retrieval may help availability, but it cannot satisfy a requirement for data the fallback does not actually provide.
B:Matching shape does not establish equivalent freshness, precision, or business meaning of the returned values.
D:One provider can offer sources with different freshness and precision. Provider identity does not establish that the alternative meets the original evidence requirement.
Question 99
An agent workflow has submitted a review request. The review service will send a signed completion callback, and no useful work remains until it arrives. Reviews can take hours. Which execution design best avoids idle model calls while allowing recovery after a process restart?
- Mark the review step complete when the service accepts the request for review.
- Persist a waiting state and callback correlation, then resume when the verified event arrives.
- Ask the model to reconsider the pending review every minute until it predicts completion.
- Keep the workflow only in a sleeping process until the callback arrives.
Explanation
Correct Answer
B:The workflow can release active execution while retaining enough durable state to match the callback to its continuation. Recovery does not require reconstructing an hours-long in-memory wait.
Incorrect Answers
A:Acceptance confirms that the review was queued, not that a decision exists. Advancing now would bypass the dependency the callback is meant to resolve.
C:Model calls cannot observe a review decision that has not been delivered. Repeated inference consumes resources without advancing the external dependency.
D:Sleeping avoids inference but leaves continuation dependent on that process surviving. The scenario requires recovery when the process restarts.
Question 100
A tool returns an authorization-denied error. Which retry policy is appropriate?
- Remove authorization headers to bypass the denial.
- Resolve the permission or identity problem before retrying the protected operation.
- Retry rapidly because all errors are transient.
- Switch to another user's credentials from prior context.
Explanation
Correct Answer
B:Authorization failures require a valid access path, not merely more attempts. The assistant should preserve the error and seek a supported remedy.
Incorrect Answers
A:Dropping credentials does not create permission and may change the failure mode.
C:An unchanged unauthorized request will generally remain unauthorized and can create unnecessary load.
D:Another identity cannot be assumed authorized for this user's task.