Question 20
An orchestrator needs consistent handoffs among agents written by different teams. What should each worker return?
- The exact same prompt originally sent to the worker
- A return contract covering status, evidence and failures
- Only a subjective confidence adjective such as excellent
- A free-form stream of internal reasoning with no completion marker
Explanation
Correct Answer
B:Structured contracts let the parent distinguish completed, partial and failed work. They also support predictable aggregation and retry decisions.
Incorrect Answers
A:A prompt does not express an execution outcome. It contributes no evidence about whether the delegated task succeeded.
C:Confidence alone does not contain findings, provenance or actionable errors. It cannot drive robust synthesis.
D:Unbounded scratch text is difficult to parse and may omit explicit outcome status. Parent orchestration needs a reliable interface.
Question 21
A coordinator notices each worker starts new searches despite a shared verified document index. What change improves efficiency without erasing responsibility?
- Provide workers a shared read-only artifact index and scoped search tasks
- Eliminate specialist agents and ask the index to draft conclusions
- Pass around unrestricted write access to all working artifacts
- Tell workers to ignore available documents and search only from memory
Explanation
Correct Answer
A:Workers can reuse validated sources while preserving separate investigations. The coordinator still tracks which evidence each conclusion used.
Incorrect Answers
B:An index stores data but does not perform the intended domain analysis and synthesis. The architecture still needs a reasoning step.
C:Shared mutable access risks races and accidental overwrites. The need is reuse of evidence, not unbounded editing privileges.
D:Unverified model recall is weaker than available primary evidence. This increases hallucination risk and duplicated search effort.
Question 22
A multi-agent audit must produce one final decision after independent policy, financial and technical checks. Which team member should own the final decision protocol?
- The technical specialist alone because code has more detail
- Any specialist that reports a pass first
- The coordinator evaluating all three specialist results
- A random worker selected after the checks finish
Explanation
Correct Answer
C:Final authority should live at an explicit aggregation point with visibility into dependencies. Specialists provide evidence within their disciplines, not unilateral global approval.
Incorrect Answers
A:Technical evidence is only one required dimension. Finance and policy constraints must also be satisfied before release.
B:One domain passing says nothing about unresolved failures elsewhere. Completion order cannot be the global release criterion.
D:Random ownership offers no predictable accountability. A documented central decision rule is needed for auditability.
Question 23
In a multi-agent review, workers repeatedly invoke each other and exhaust the budget without producing a summary. Which orchestration change is most appropriate?
- Ask workers to use more creative language in their prompts
- Acyclic delegation with a coordinator-owned stopping rule
- Permit each worker to spawn unlimited peers to find an answer
- Collect a bigger transcript before deciding whether to stop
Explanation
Correct Answer
B:Clear parent-to-worker responsibilities and finite fan-out prevent cycles. A completion contract lets the parent decide when enough evidence exists.
Incorrect Answers
A:Style changes do not enforce acyclic task ownership or a hard budget. The flaw is in control flow.
C:Additional recursive delegation can worsen circular dependencies and cost. It does not establish who decides termination.
D:More history records the failure but does not prevent it. The orchestrator needs a bounded work graph and exit criteria.
Question 24
A customer support coordinator receives an account-check result and a policy-check result that disagree on eligibility. The service has a published precedence rule. What should the coordinator do?
- Suppress the disagreement in the audit trail
- Apply the published precedence rule and preserve both observations
- Ask both workers to independently apply the refund again
- Select the answer with the longest justification
Explanation
Correct Answer
B:The central decision point can use a deterministic policy precedence without relying on worker preference. Retaining evidence explains the resolution.
Incorrect Answers
A:Hidden conflicts prevent later investigation and can conceal policy defects. The coordinator should retain both inputs and the resolution basis.
C:Replaying a side effect does not resolve conflicting eligibility evidence. It can create duplicate transactions.
D:Explanation length is not a reliable tie-breaker when there is an explicit business policy. Authority comes from the rule and evidence.
Question 25
A lead research agent needs a short independent search on a narrow topic. How should it provide the task to a new subagent?
- Pass every private note and unrelated prior conversation to the subagent
- Spawn the worker with no goal and infer its purpose afterward
- Grant all available write tools before determining the task
- Supply a specific objective, boundaries and expected return format
Explanation
Correct Answer
D:A bounded task gives the specialist enough direction without importing unrelated parent history. A concise result contract supports later synthesis.
Incorrect Answers
A:Overbroad context wastes tokens and may expose sensitive information. The worker only needs material relevant to its assigned task.
B:Subagents cannot reliably choose the parent's intended scope from an empty instruction. Ambiguity produces off-target work.
C:Capabilities should follow the requested role. Unnecessary write access increases operational and security risk.
Question 26
A code investigator must inspect a repository but must not modify files. What subagent configuration most directly supports the boundary?
- Give the subagent shell and edit access with a polite no-edit request
- Ask the parent to erase any edits after they happen
- Assign only read-oriented tools and explicit inspection instructions
- Let the child inherit every parent tool for convenience
Explanation
Correct Answer
C:Tool exposure is an enforcement layer beyond prompt guidance. Without mutation tools, the investigator cannot casually edit repository contents.
Incorrect Answers
A:A prohibition in prose is not as strong as withholding write capabilities. The request could still be violated by an erroneous tool selection.
B:Rollback is less reliable than preventing side effects. Some changes may escape detection or affect external systems.
D:Unfiltered inheritance can expose write or administrative operations irrelevant to investigation. Least privilege reduces the blast radius.
Question 27
A coordinator wants three independent hypotheses about why a test suite fails. Why might separate subagent contexts be helpful?
- Separate contexts allow subagents to bypass parent permissions automatically
- Separate contexts guarantee that every hypothesis will be correct
- Each worker explores its own hypothesis in isolated context
- Separate contexts eliminate the need to report findings to the parent
Explanation
Correct Answer
C:Isolation reduces shared speculation and lets the parent compare findings. The final conclusion can draw on evidence instead of a single contaminated thread.
Incorrect Answers
A:Context separation is not a privilege escalation mechanism. Tool scopes and authorization remain independently controlled.
B:Isolation changes information flow, not accuracy. Each result still requires verification against the actual tests and code.
D:The parent still needs results to resolve the incident. Independent workers cannot replace the aggregation stage.
Question 28
A subagent finishes a large code search but returns every scanned file to the parent. What should the team change at the task boundary?
- Replay the search in the parent to double token usage
- Ask for decisive locations, evidence snippets and relevant file paths
- Give the child unrestricted permission to rewrite files instead
- Force the subagent to omit all file locations
Explanation
Correct Answer
B:A purpose-built handoff compresses the investigation while leaving navigable references. The parent can request deeper details only when needed.
Incorrect Answers
A:Re-executing work ignores the efficiency of delegated exploration. It also risks inconsistent search results when files change.
C:Editing privileges do not address a verbose read-only result. The fault is output scoping.
D:Without provenance the parent cannot verify or continue work. Compression should preserve key pointers, not eliminate evidence.
Question 29
A financial-data worker needs access to the pricing lookup but not the refund mutation endpoint. Which distribution strategy fits?
- Give the worker pricing lookup, not refund authorization
- Make every worker the refund administrator to simplify routing
- Expose both tools and rely on the worker's memory to avoid refunds
- Hide pricing access and ask the model to estimate current prices
Explanation
Correct Answer
A:Task-specific tool lists reduce accidental sensitive operations. The parent should still validate data before asking an authorized actor to apply a refund.
Incorrect Answers
B:Uniform broad access increases risk without improving the read-only pricing workflow.
C:Prompt-only boundaries are weaker than limiting capabilities. The pricing task does not justify a financial write tool.
D:An inference cannot replace a live authorized lookup when accuracy is required. The worker needs the read tool, not the mutation tool.
Question 30
A parent agent asks a child to draft an executive brief from verified research notes. Which data should accompany the delegation?
- Every source document in full regardless of relevance
- The parent company's secret credentials with the research notes
- Only an instruction to be persuasive with no source material
- The validated findings, target audience and citation references
Explanation
Correct Answer
D:The worker can write within scope using evidence and output expectations. It need not receive the entire underlying orchestration transcript.
Incorrect Answers
A:Indiscriminate context transfer can crowd out useful instructions. Supplying curated evidence is more reliable and efficient.
B:Credentials do not help draft an authorized brief. Passing secrets needlessly enlarges the exposure surface.
C:Style guidance cannot substitute for factual inputs. An unsupported brief may invent conclusions and lose auditability.