Anthropic CCDV-F (Claude Certified Developer - Foundations) Practice Test 2

Question 10

A customer workflow requires a strict maximum of six model-tool cycles per request. Where should the limit be implemented?

  1. Name a supervisor responsible for delegation and termination.
  2. For this workflow, avoid agent autonomy when a predetermined sequence already meets the requirements.
  3. Create an explicit draft-critique workflow with a defined exit condition.
  4. Enforce a loop iteration budget in the orchestration code.

Explanation

Correct Answer

D:

Deterministic harness limits end execution independently of model intent; an instruction to stop is not a reliable execution boundary.

Incorrect Answers

A:

A supervisor can track completion criteria and prevent uncontrolled peer handoffs; the workflow has one accountable termination decision; testing bounded tool-loop steps would still fail after improving controlled delegation ownership.

B:

An agent introduces nondeterministic planning and extra testing surface without improving a process whose branches can be specified directly; for bounded tool-loop steps, a control centered on unnecessary autonomous decisions is not the appropriate substitute.

C:

A known two-stage procedure is easier to test as a fixed workflow than as an unrestricted multi-agent conversation; a remedy for bounded critique stage is not a substitute for bounded tool-loop steps.

Question 11

An agent working on a large repository loses track of its main objective after several explorations. How can architecture preserve task focus?

  1. Give exploratory subtasks separate context and feed bounded summaries to the coordinator.
  2. Enable the model to choose a sequence of diagnostic tools under bounded permissions.
  3. Put in place a deterministic workflow with explicit stage transitions.
  4. Locate a deterministic approval gate between planning and the booking tool.

Explanation

Correct Answer

A:

Isolating exploratory traces protects the main context from unrelated evidence while retaining task-relevant conclusions.

Incorrect Answers

B:

Because observations determine subsequent steps, a controlled agent can adapt its plan while the harness constrains its scope and cost; the consequence is improved exploratory diagnostic sequence, not verified preserving coordinator focus.

C:

A fixed workflow makes ordering, validation, and retry boundaries visible in application code; an autonomous planner would add choices the requirement does not need; for this request, preserving coordinator focus is the missing piece, not guaranteed processing order.

D:

The agent may draft an itinerary, but a separate permission check must decide whether the consequential booking invocation is authorized; it follows user-approved irreversible action constraints but leaves preserving coordinator focus unspecified.

Question 12

Two specialized assistants alternately hand work to each other without a clear owner, causing loops. Which pattern adds control?

  1. Route bounded research to subagents and return concise results to the parent.
  2. Appoint a supervisor responsible for delegation and termination.
  3. Review the refund threshold in ordinary application logic.
  4. Implement a loop iteration budget in the orchestration code.

Explanation

Correct Answer

B:

A supervisor can track completion criteria and prevent uncontrolled peer handoffs; the workflow has one accountable termination decision.

Incorrect Answers

A:

Separate subagent contexts retain intermediate tool output locally, so the coordinator receives only the information needed to synthesize; this implements isolation of verbose research rather than the demanded controlled delegation ownership.

C:

Monetary policy is deterministic and should be enforced outside natural-language generation; the model can supply context but cannot override the branch; where controlled delegation ownership matters, a control for deterministic business policy leaves that test unmet.

D:

Deterministic harness limits end execution independently of model intent; an instruction to stop is not a reliable execution boundary; changing bounded tool-loop steps does not establish controlled delegation ownership in this application.

Question 13

A billing service must guarantee that refunds above a threshold go to a human, regardless of what Claude recommends. How should the branch be chosen?

  1. Evaluate the refund threshold in ordinary application logic.
  2. For this workflow, maintain structured working state for completed steps and observations.
  3. Employ an agent loop that chooses tools from the current evidence.
  4. As an alternative, pass purpose-limited task briefs and data subsets to each worker.
  5. Construct an explicit draft-critique workflow with a defined exit condition.

Explanation

Correct Answer

A:

Monetary policy is deterministic and should be enforced outside natural-language generation; the model can supply context but cannot override the branch.

Incorrect Answers

B:

A durable task-state record lets orchestration distinguish new actions from already performed ones without depending on an ever-growing transcript; a remedy for tracking completed actions is not a substitute for deterministic business policy.

C:

The route is not known in advance; a model-driven loop can evaluate observations and choose the next authorized step, while the harness enforces limits; the scenario's controlling requirement is deterministic business policy, not the related dynamic next-action selection concern.

D:

Scoped delegation reduces both context noise and unnecessary disclosure, while the manager retains responsibility for the combined result; this may help need-to-know subagent inputs but cannot by itself deliver deterministic business policy.

E:

A known two-stage procedure is easier to test as a fixed workflow than as an unrestricted multi-agent conversation; for this request, deterministic business policy is the missing piece, not bounded critique stage.

Question 14

A troubleshooting agent receives a vague report of “the dashboard is broken,” but expensive diagnostic tools require the affected environment and time window. What is the best first step?

  1. Give the assistant permanent administrator permissions to make the investigation quicker.
  2. Ask for the environment and time window before invoking costly, broad diagnostics.
  3. Assume the production environment and the past 24 hours without telling the user.
  4. Start every available diagnostic tool in parallel across all environments.

Explanation

Correct Answer

B:

The task lacks the inputs needed to target a useful investigation. A clarifying question reduces wasted searches and prevents indiscriminate access while preserving the ability to use an adaptive tool loop after the scope is known.

Incorrect Answers

A:

More privileges do not resolve an ambiguous request. The execution layer should grant only task-appropriate read permissions, especially before the user identifies the affected system.

C:

Defaulting silently to a high-impact environment risks irrelevant evidence and mistaken conclusions. The agent should obtain or explicitly confirm critical scope rather than invent it.

D:

Maximum tool fan-out spends resources and may touch systems outside the user’s intended scope. It does not compensate for missing environmental boundaries or an undefined incident window.

Question 15

A report generator needs one model to draft and a separate model to critique against a rubric, then stops. Which design is adequate?

  1. Build an explicit draft-critique workflow with a defined exit condition.
  2. Compare evidence and uncertainty before producing its final synthesis.
  3. Install a supervisor responsible for delegation and termination.
  4. Test the refund threshold in ordinary application logic.

Explanation

Correct Answer

A:

A known two-stage procedure is easier to test as a fixed workflow than as an unrestricted multi-agent conversation.

Incorrect Answers

B:

Parallel delegation accelerates collection, but agreement does not arise automatically; the coordinator must evaluate conflicts and decide whether to request more evidence; for this request, bounded critique stage is the missing piece, not evidence reconciliation.

C:

A supervisor can track completion criteria and prevent uncontrolled peer handoffs; the workflow has one accountable termination decision; this may help controlled delegation ownership but cannot by itself deliver bounded critique stage.

D:

Monetary policy is deterministic and should be enforced outside natural-language generation; the model can supply context but cannot override the branch; it follows deterministic business policy constraints but leaves bounded critique stage unspecified.

Question 16

An agent repeatedly rechecks the same source because it has no record of successful actions. What addition best addresses this?

  1. Maintain structured working state for completed steps and observations.
  2. At the application layer, fan out specialized reviews and aggregate their findings in a coordinator.
  3. Put in place a loop iteration budget in the orchestration code.
  4. For this workflow, pass purpose-limited task briefs and data subsets to each worker.

Explanation

Correct Answer

A:

A durable task-state record lets orchestration distinguish new actions from already performed ones without depending on an ever-growing transcript.

Incorrect Answers

B:

Parallel specialists can inspect distinct criteria independently; the parent reconciles results and applies a final decision policy; for tracking completed actions, a control centered on independent parallel reviews is not the appropriate substitute.

C:

Deterministic harness limits end execution independently of model intent; an instruction to stop is not a reliable execution boundary; when the goal is tracking completed actions, optimizing bounded tool-loop steps misses that operational boundary.

D:

Scoped delegation reduces both context noise and unnecessary disclosure, while the manager retains responsibility for the combined result; changing need-to-know subagent inputs does not establish tracking completed actions in this application.

Question 17

A manager agent is being redesigned to lower token use and enforce safe execution. Which TWO changes directly achieve those separate objectives? Choose TWO.

  1. Return short, source-linked summaries from specialized workers.
  2. Give every worker the entire raw chat transcript.
  3. Enforce tool permissions in the execution layer.
  4. Ask the manager to promise not to exceed a limit.
  5. Increase the maximum response length for each worker.

Explanation

Correct Answers

A:

Subagent summaries prevent irrelevant intermediate observations from filling the parent context while retaining essential provenance. The coordinator can still cite individual sources without inheriting full diagnostic logs from every worker.

C:

Runtime permission checks prevent unsafe actions independently of the model’s stated intent or confidence. The actual dispatcher must decide whether an identity is permitted to call each named operation.

Incorrect Answers

B:

Unfiltered history duplicates context cost and can leak unrelated information; it does not achieve lean delegation. Copying unrelated history into every child also increases the chance that sensitive context reaches the wrong role.

D:

Natural-language promises are not an execution boundary; the harness must enforce hard constraints. A budget written only in the prompt cannot terminate an already-running unauthorized shell or network operation.

E:

More output tokens enlarge intermediate reports rather than reduce context consumption or guarantee tool safety. A larger completion allowance increases potential output volume while leaving tool authorization unchanged.

Question 18

A manager must ensure each specialist receives only a relevant slice of a sensitive case file. What coordination approach is suitable?

  1. Implement an explicit draft-critique workflow with a defined exit condition.
  2. Insist on a concrete benefit from independent context or genuinely parallel subproblems.
  3. At the application layer, maintain structured working state for completed steps and observations.
  4. Pass purpose-limited task briefs and data subsets to each worker.

Explanation

Correct Answer

D:

Scoped delegation reduces both context noise and unnecessary disclosure, while the manager retains responsibility for the combined result.

Incorrect Answers

A:

A known two-stage procedure is easier to test as a fixed workflow than as an unrestricted multi-agent conversation; where need-to-know subagent inputs matters, a control for bounded critique stage leaves that test unmet.

B:

Extra agents add coordination, latency, and failure modes; they earn their cost only when workload structure provides useful separation; the scenario's controlling requirement is need-to-know subagent inputs, not the related multi-criterion integration concern.

C:

A durable task-state record lets orchestration distinguish new actions from already performed ones without depending on an ever-growing transcript; the effect concerns tracking completed actions; the success condition is need-to-know subagent inputs.

Question 19

A custom harness receives a Claude response whose stop_reason is tool_use. What must the application do before asking Claude for a final answer?

  1. Create a custom Messages API harness that owns dispatch and state.
  2. Store job state outside ephemeral request workers and resume from checkpoints.
  3. Execute the requested client tools and return linked tool_result blocks in a new user message.
  4. Configure deterministic hooks around the tool lifecycle.
  5. Prefer an execution environment with authorized network access to the private tool.

Explanation

Correct Answer

C:

Client tools do not run merely because the model emitted a tool_use block; the caller must dispatch them and preserve the tool_use identifiers.

Incorrect Answers

A:

Direct orchestration exposes scheduling, persistence, error handling, and tool-execution boundaries; it trades framework convenience for precise integration control; this targets custom orchestrator control, while the system must demonstrate handling a tool request.

B:

Long-running tasks must survive worker termination; relying on a single request's memory risks losing progress and repeating side effects; it changes durable long-running agent execution; it does not prove handling a tool request in this case.

D:

Hook handlers are executed by the framework at defined events and can record metadata independently of model-generated prose; the scenario's controlling requirement is handling a tool request, not the related deterministic tool event instrumentation concern.

E:

Deployment placement determines reachability and data boundaries; neither agent quality nor an SDK name can bypass those infrastructure constraints; this implements private-network deployment fit rather than the demanded handling a tool request.

Question 20

A service receives an end_turn response after one tool result. How should its dispatch loop react?

  1. Specify and validate a typed input schema at the tool boundary.
  2. Interpret content blocks by type rather than treating the response as one string.
  3. Send a structured tool_result describing the failure without fabricating inventory.
  4. Exit the tool loop and deliver the completed assistant response.

Explanation

Correct Answer

D:

end_turn indicates the current assistant turn has finished, so blindly issuing another tool call would add unnecessary steps.

Incorrect Answers

A:

Explicit fields and runtime validation reject malformed inputs before touching external systems; model-generated arguments are not trusted by default; the recognizing completion criterion differs from tool argument validation; the gap persists.

B:

Tool_use blocks carry structured inputs and IDs, not user-facing prose; the harness must branch on content type and stop reason; where recognizing completion matters, a control for content-block-aware parsing leaves that test unmet.

C:

Reporting the real error lets the model revise its plan or explain the limitation; silent invention hides a failed integration; this may help surfacing client-tool errors but cannot by itself deliver recognizing completion.

Claude Certified Developer - Foundations (CCDV-F) Practice Tests