CertifyAIxAce every AI certification.

CCDV-F sample questions with answers

10 free practice questions for the Claude Certified Developer: Foundations exam. Try each one, then open the answer to see why the right option wins and every other option loses.

Question 1Agents and Workflows

Jessop Analytics runs an agent for days-long investigations. It already uses server-side compaction to keep the active context small, but a few key findings, such as confirmed root causes, must survive every summarisation intact. Which addition does Anthropic's documentation recommend?

  1. A.

    Replace compaction with context editing, so that the key findings can never be cleared from context

  2. B.

    Keep the findings only in the conversation, since compaction preserves everything the model marks important

  3. C.

    Compact more often, so that each summary covers less history and loses fewer details as a result

  4. D.

    Pair compaction with the memory tool, so the findings that must survive are written to memory files

Show answer

Answer: D

The memory tool documentation recommends using both: compaction keeps context small, memory preserves what must survive summarisation.

  • A. Incorrect: context editing clears tool results; it is not a mechanism for protecting findings the agent must keep.
  • B. Incorrect: nothing guarantees a summary keeps specific items, and the guidance warns compaction can lose them.
  • C. Incorrect: more frequent compaction means more summarisation events, each of which can drop detail.
  • D. Correct: memory makes the critical findings durable outside the window while compaction keeps the working context small.
Question 2Agents and Workflows

Kinsale Software's Agent SDK integration reads the result field of the final result message and stores it as the agent's answer. Most runs work, but some store an empty value and the pipeline then publishes a blank report. The failing runs all hit the configured turn limit. What is the cause and the fix?

  1. A.

    The result field exists only on success, so check the result subtype before reading the output

  2. B.

    The session was compacted before the limit, so the answer text was removed during summarisation

  3. C.

    The turn limit truncates the final answer, so raise the output token limit for each response

  4. D.

    The SDK drops result messages when a limit is hit, so read the last assistant message instead

Show answer

Answer: A

When a run hits a limit, the result message carries an error subtype and no result field, so the subtype must be checked first.

  • A. Correct: only the success subtype includes result, so the code must check the subtype before using the output.
  • B. Incorrect: compaction is unrelated to the turn cap, and the failures correlate exactly with hitting the limit.
  • C. Incorrect: output token limits affect response length; the missing field comes from the turn-limit error subtype.
  • D. Incorrect: a result message is still sent when a limit is hit; it simply carries an error subtype.
Question 3Agents and Workflows

Idris Fintech's Agent SDK agent edits Python files. The team's rule is that the code formatter must run after every file edit, every time. Adding 'always run the formatter after editing' to the system prompt works most of the time but occasionally gets skipped. Which construction approach makes the action deterministic?

  1. A.

    Repeat the formatting instruction at both the start and the end of the system prompt

  2. B.

    Register a PostToolUse hook for the edit tools that runs the formatter after each edit completes

  3. C.

    Raise the effort level so the agent reasons more carefully about following its instructions

  4. D.

    Give the agent a dedicated format tool and describe clearly in its description when to call it

Show answer

Answer: B

A hook runs your code at a fixed point in the loop every time, which a prompt instruction cannot guarantee.

  • A. Incorrect: repetition may raise compliance, but it still leaves the formatting step as a decision the model makes.
  • B. Correct: a PostToolUse hook runs the formatter after every matching edit, independent of the model's choices.
  • C. Incorrect: more reasoning effort does not turn an instruction into a guaranteed action at a fixed point in the loop.
  • D. Incorrect: a tool must still be chosen by the model, so a missed call remains possible however good the description.
Question 4Agents and Workflows

Lorimer Telecom's support agent holds long, conversational troubleshooting sessions with customers, with extensive back-and-forth over an hour or more, and occasionally approaches the context limit. The conversation's flow matters more than recalling exact early details. Which context-management pattern does Anthropic's guidance match to this kind of task?

  1. A.

    Structured note-taking only, with every customer message written to a file and read back later

  2. B.

    Compaction, summarising the conversation near the limit and continuing with the summary

  3. C.

    Restarting the conversation from scratch whenever the context approaches its limit

  4. D.

    A multi-agent architecture, with parallel subagents exploring separate parts of the conversation

Show answer

Answer: B

Anthropic says compaction maintains conversational flow for tasks that need extensive back-and-forth.

  • A. Incorrect: recording every message suits milestone-driven work and adds overhead to a conversational task.
  • B. Correct: compaction keeps a long back-and-forth conversation going by carrying a summary into a fresh context.
  • C. Incorrect: restarting discards the conversation, which is the opposite of preserving its flow.
  • D. Incorrect: parallel subagents suit breadth-first exploration, not a single conversational thread with a customer.
Question 5Agents and Workflows

Oakhurst Data's engineers are deciding whether their Claude agent should delegate some tasks to subagents. They want a clear statement of what subagents actually provide before they commit to the design. Which TWO statements accurately describe benefits of subagents? (Select TWO.)

Choose 2.

  1. A.

    Intermediate tool calls stay in the subagent, and only its final message returns to the parent

  2. B.

    Independent subtasks can run concurrently, so they finish in about the time of the slowest one

  3. C.

    Each subagent starts with the parent's full conversation history, so no context has to be passed

  4. D.

    Delegating to subagents reduces the total number of tokens a task consumes compared with one agent

  5. E.

    Subagents make an agent's results deterministic, because each one follows a fixed code path

Show answer

Answer: A, B

Subagents provide context isolation and parallel execution; they do not inherit history, save tokens overall, or make runs deterministic.

  • A. Correct: context isolation means the parent receives only the subagent's final message, not its whole working transcript.
  • B. Correct: concurrent execution lets independent subtasks finish in roughly the time of the slowest one.
  • C. Incorrect: a non-fork subagent starts fresh; only the delegation prompt carries information from the parent.
  • D. Incorrect: each subagent makes its own model requests, so total token use usually rises rather than falls.
  • E. Incorrect: each subagent is its own model-driven loop, so delegation does not make results deterministic.
Question 6Agents and Workflows

Calloway Bank's assistant must answer customer questions while also screening each request for attempts to extract other customers' data. In testing, a single prompt asked to do both answers well but misses many screening cases. Latency matters, so the two concerns should not run one after the other. Which design follows Anthropic's workflow guidance?

  1. A.

    Voting: run the combined answer-and-screen prompt several times and keep the most common output

  2. B.

    Prompt chaining: screen the request first, then pass it to the answering prompt once it is cleared

  3. C.

    Sectioning: one model instance answers while a separate instance screens the same request in parallel

  4. D.

    Routing: classify each request as risky or safe, and send only the risky requests to screening

Show answer

Answer: C

Separate parallel calls for the core answer and the guardrail are Anthropic's sectioning example.

  • A. Incorrect: voting repeats the combined prompt that already misses screening cases, so it inherits the same weakness.
  • B. Incorrect: chaining screen-then-answer is safe but sequential, which breaks the stated latency requirement.
  • C. Correct: separate parallel calls give each concern focused attention without adding sequential latency.
  • D. Incorrect: routing needs the risky-or-safe decision made up front, which is the screening job itself.
Question 7Agents and Workflows

Tamsin Analytics' lead agent delegates market research with instructions such as 'research the battery supply chain'. Reviews find that two subagents ran nearly identical searches, one strayed into a different decade, and none returned findings in a form the lead could merge easily. What change to the lead's delegation would address all three problems?

  1. A.

    Give every subagent the lead's full conversation history so each knows the others' work

  2. B.

    Increase the number of subagents so that gaps left by off-topic subagents are covered by others

  3. C.

    Give each subagent an objective, an output format, guidance on tools and sources, and clear task boundaries

  4. D.

    Replace the subagents with one agent that runs all searches in sequence so nothing is duplicated

Show answer

Answer: C

Anthropic found that each subagent needs an objective, an output format, tool and source guidance, and clear task boundaries.

  • A. Incorrect: sharing full history undermines context isolation and still leaves each subagent's task undefined.
  • B. Incorrect: more subagents with equally vague briefs multiply the duplication instead of dividing the work.
  • C. Correct: a precise objective, boundaries and output format address the duplication, the drift and the merge problem together.
  • D. Incorrect: one sequential agent removes duplication by giving up the parallel breadth that justified the subagents in the first place.
Question 8Agents and Workflows

Northgate Bank registered two PreToolUse hooks for every tool call. Hook A looks up the caller's tenant and stores it in a shared variable; hook B reads that variable to decide whether to allow the call. Most calls work, but some are allowed or denied using the previous call's tenant. What is the cause?

  1. A.

    Hooks run in the model's context window, so the variable is lost at compaction

  2. B.

    Only one PreToolUse hook is allowed per event, so the second replaces the first

  3. C.

    PreToolUse hooks run after the tool executes, so B sees the previous call's tenant

  4. D.

    Matching hooks for an event run in parallel, so B cannot rely on A having run first

Show answer

Answer: D

When an event fires, all matching hooks run in parallel with non-deterministic completion order, so hooks must not depend on each other.

  • A. Incorrect: hooks run in the application process, not in the context window, so compaction cannot affect them.
  • B. Incorrect: several hooks per event are supported; the documentation shows three independent checks on one event.
  • C. Incorrect: PreToolUse fires before the tool runs, which is why it can block or modify the call.
  • D. Correct: parallel execution with non-deterministic completion order means B can run before A updates the tenant.
Question 9Agents and Workflows

Mallory Retail's Managed Agents session uses the built-in agent tools and a connected inventory MCP server. Built-in tools such as bash run without interruption, but every call to an inventory MCP tool pauses the session until someone responds. The team trusts this MCP server fully. What explains the pauses, and what is the fix?

  1. A.

    The MCP server is unreachable, so each call waits for a network timeout before it is retried

  2. B.

    MCP toolsets default to alwaysask; set the trusted server's toolset policy to alwaysallow

  3. C.

    The agent's model does not support MCP tools, so each call falls back to a slower manual approval path

  4. D.

    Built-in tools and MCP tools cannot run in the same session, so the MCP calls are queued behind them

Show answer

Answer: B

Managed Agents applies alwaysallow to the agent toolset by default but alwaysask to MCP toolsets, which pauses for approval.

  • A. Incorrect: an unreachable server produces errors or timeouts, not a session pausing for approval.
  • B. Correct: the MCP toolset's default policy asks for approval, and a trusted server can be set to always_allow.
  • C. Incorrect: no model-dependent fallback to approval exists; the MCP toolset's default policy causes the pause.
  • D. Incorrect: sessions can combine the agent toolset with MCP toolsets; the pause comes from the permission policy.
Question 10Agents and Workflows

Aurelian Press translates literary essays with Claude. First drafts are accurate but miss tone and idiom, and when an editor writes specific critiques, a revised draft is clearly better. Tests show Claude can also produce useful critiques of its own drafts against a written style rubric. Which workflow pattern is the best fit?

  1. A.

    Orchestrator-workers, where a central model assigns each sentence to a separate worker model

  2. B.

    Parallelization by sectioning, translating each paragraph in its own call and joining the results

  3. C.

    Routing, where each essay is sent to a translation prompt specialised for its literary genre

  4. D.

    Evaluator-optimizer, where one call drafts and another evaluates against the rubric in a loop

Show answer

Answer: D

Clear criteria, feedback that demonstrably improves drafts, and a model able to give that feedback are the two signs of good fit for evaluator-optimizer.

  • A. Incorrect: sentence-level delegation fragments the text and answers an unpredictable-subtask problem that does not exist here.
  • B. Incorrect: paragraph sectioning breaks tone across the essay, which is the exact quality that is already weak.
  • C. Incorrect: genre routing still yields one pass per essay and never feeds critique back into a revised draft.
  • D. Correct: the rubric gives clear criteria, and demonstrated gains from feedback make an evaluate-and-revise loop worth its cost.

Keep going with 624 more CCDV-F questions

Free papers every day, in the real exam formats, with progress by exam domain. Unlock every paper and timed mock exam when you are ready.

Practise freeSee all CCDV-F papers
CCDV-F sample questions with answers (10 free) · CertifyAIx