Mock Exam 2

60 items · 120 minutes · four scenarios

Scenarios in this paper: 4 (Developer Productivity), 5 (Claude Code for CI), 1 (Customer Support — second visit), 3 (Multi-Agent Research — second visit). Together with Mock 1, the two papers cover all six scenarios in the bank.

This paper is deliberately not blueprint-weighted. Its actual distribution is D1 22 · D2 15 · D3 6 · D4 9 · D5 8 — because the four scenarios it draws on are the agentic and tool-heavy ones, exactly as the guide maps them. Mock 1 is the blueprint-weighted simulation; use that one to predict your score. Use this one to stress the two domains that together are 45% of the real exam, and to see all six scenarios before exam day.

It is also harder than Mock 1: fewer pure-recall items, more items where two options are defensible and you must pick the better one. Expect a lower raw score here, and don't read that as a regression.

Instructions

Answer every item. No penalty for a wrong answer.
Each item states how many options to select. Unless stated, select one.
Target pace: 30 minutes per scenario.
Multiple-response items are all-or-nothing. Partial credit does not exist.

Answers: mock-exam-2-answer-key.md. Scoring: scoring-grid.md.

Scenario A — Developer Productivity with Claude

A 200-engineer organization is deploying an internal agent built on the Claude Agent SDK to help developers navigate a large legacy monorepo, write tests, and update documentation. It uses the built-in Read, Write, Edit, Bash, Grep, and Glob tools, plus MCP servers for the internal service catalogue and the ticketing system. Adoption is uneven and some teams report the agent "makes things up" about the codebase.

1

A developer asks the agent to explain how billing reconciliation works in an unfamiliar 200-file service. Which approach is most efficient?

A. Read every file to build a complete picture.
B. Grep for entry points, then Read selectively to follow the imports.
C. Glob for all source files and Read the twenty largest.
D. Ask the agent to describe how services of this kind are typically structured.

2

Late in a long refactoring session the agent stops naming specific classes and begins describing "the typical pattern for this kind of service." Its suggestions no longer match the code. Diagnosis and fix?

A. Context degradation; delegate exploration to subagents and keep a scratchpad file of established findings.
B. Insufficient model capability; enable extended thinking.
C. An unclear goal; restate the refactoring objective in more detail.
D. The codebase is too large for any agent; split it across repositories.

3

Which built-in tool finds every file whose path matches **/*.integration.ts?

A. Glob
B. Grep
C. Read
D. Bash

4

The agent must change a config value that appears once in a file with a distinctive surrounding line. Which tool?

A. Edit, anchored on the unique surrounding text.
B. Read then Write.
C. Bash with sed.
D. Grep then Write.

5

Every subagent in the system is given all fourteen tools "for flexibility." Which two problems does this create? (Select two.)

A. Reduced tool-selection accuracy within each subagent.
B. Violation of least privilege — subagents can take actions outside their role.
C. Subagents begin communicating directly with each other.
D. Higher latency during tool discovery at connection time.

6

An MCP-provided search_service_catalogue tool is consistently ignored in favour of the built-in Grep, even where it would be better. Fix?

A. Enhance the MCP tool's description to state explicitly what it does that Grep cannot.
B. Remove Grep from the agent's allowed tools.
C. Mark the MCP tool as preferred in .mcp.json.
D. Register the MCP tool twice so it appears more prominent.

7

The team needs Jira integration. What does the exam prefer?

A. Use an existing community MCP server for Jira.
B. Build a custom MCP server wrapping the Jira REST API.
C. Give the agent Bash access and let it call curl.
D. Deploy a self-hosted MCP gateway in front of Jira.

8

When are an MCP server's tools discovered, and how are they available afterwards?

A. At connection time; all discovered tools are then simultaneously available.
B. On first use; each must be enabled individually.
C. Per request, based on the prompt's content.
D. At connection time, but only the first ten load.

9

The internal service catalogue exposes a large browsable set of service definitions the agent should be able to consult. Which MCP primitive fits?

A. Resources.
B. A tool per service.
C. A single get_service tool taking a name.
D. Embedding the catalogue in CLAUDE.md.

10

A get_service_owner tool returns null with isError: false both when a service has no registered owner and when the caller lacks directory permission. What is wrong?

A. Nothing; null correctly represents "no owner found."
B. An access failure is being reported as a valid empty result; permission denial needs isError: true with a permission category.
C. The tool should retry with elevated credentials.
D. The agent should treat all null returns as errors.

11

The organization wants the agent to update documentation but never to modify production Terraform. Which mechanism is correct?

A. Scope the agent's tools and add a PreToolUse gate blocking writes under the Terraform paths.
B. Add a CLAUDE.md instruction never to modify Terraform.
C. Add a prominent warning to the agent's system prompt.
D. Ask developers to review every diff before applying it.

12

A single analyze_code tool is used for finding dead code, summarizing modules, and checking test coverage. Quality is poor across all three. Best fix?

A. Add detailed instructions to the description covering all three use cases.
B. Split it into three purpose-scoped tools.
C. Add a mode parameter with three allowed values.
D. Raise the tool's output token limit.

13

Analysis shows 85% of analyze_code calls are one narrow task: checking whether a module has test coverage. The other 15% vary. Best design?

A. Add a scoped check_test_coverage tool for the common case and keep the general tool for the rest.
B. Keep one general tool and improve its description.
C. Replace the general tool entirely with the scoped one.
D. Add a task_type enum to the general tool.

14

Which elements belong in a tool description to support correct selection? (Select three.)

A. Accepted input formats and required prefixes.
B. Example queries the tool is appropriate for.
C. Explicit boundaries — what it does not handle, and which tool does.
D. Average latency and cost per call.

15

The agent has a well-described lookup_runbook tool but keeps using web_search. Descriptions are clear and non-overlapping. What should be audited next?

A. The system prompt, for keyword-sensitive instructions steering toward search.
B. lookup_runbook's output schema.
C. Whether lookup_runbook appears before web_search in the tools array.
D. The temperature setting.

Scenario B — Claude Code for Continuous Integration

A team has integrated Claude Code into its CI pipeline for automated code review, test generation, and pull request feedback. Early results produced so many low-value comments that developers began ignoring the bot entirely. The team must minimize false positives while keeping useful coverage, and the pre-merge review must not delay merges.

16

The current review instruction is "flag outdated comments." Many correct comments are flagged. Best replacement?

A. "Flag comments only when the claimed behaviour contradicts the actual code behaviour."
B. "Be conservative when flagging comments."
C. "Only flag comments you are highly confident are outdated."
D. "Flag comments that appear older than the surrounding code."

17

Why does the exam treat false positives as more damaging than false negatives here?

A. They consume more tokens per finding.
B. They destroy developer trust, after which true positives are ignored too.
C. They fail the pipeline and block merges.
D. They are harder to measure than false negatives.

18

70% of the reviewer's false positives come from one category: "possible null dereference." Which two responses are appropriate? (Select two.)

A. Temporarily disable that category while its criteria are rewritten.
B. Rewrite the category's criteria with explicit qualifying conditions and counter-examples.
C. Instruct the model to be more careful with null-dereference findings.
D. Lower the category from critical to informational and keep reporting it.

19

What makes severity criteria operational rather than aspirational?

A. Concrete code examples at each level, including explicit "NOT critical" counter-examples.
B. A numeric rubric the model applies to each finding.
C. An instruction to match the issue tracker's severity conventions.
D. A required confidence score alongside each severity.

20

Which schema field best supports analysing why the reviewer produced a given finding?

A. detected_pattern
B. confidence
C. severity
D. line_number

21

Which flags give a non-interactive run with machine-parseable, schema-constrained output? (Select three.)

A. -p / --print
B. --output-format json
C. --json-schema
D. --non-interactive

22

The CI review job runs in the same session that generated the code, immediately after generation. It rarely finds real problems. Why?

A. The generating session carries the assumptions that produced any errors; review needs a fresh, independent context.
B. The review prompt needs to be longer and more specific.
C. Extended thinking should be enabled for the review step.
D. The review should cover the full repository rather than the diff.

23

Which review configuration does the exam rank highest?

A. A separate reviewer instance with fresh context.
B. The generating instance reviewing its own output.
C. The generating instance with extended thinking enabled.
D. Two sequential self-review passes.

24

Which CI workload suits the Message Batches API?

A. The nightly aggregate quality report across all repositories.
B. The pre-merge review that gates pull requests.
C. Both, since both are automated.
D. Neither, since batch does not support tools.

25

How does the CI run pick up the team's conventions?

A. From CLAUDE.md, read in headless mode too.
B. Conventions must be inlined in the CI prompt; memory files don't load headlessly.
C. Via a --rules flag.
D. Via CI environment variables.

26

The team wants generated tests to follow a specific arrange/act/assert shape. What steers this most effectively?

A. 2–3 concrete input/output examples of tests in the desired shape.
B. A detailed prose specification of the shape.
C. A single canonical example.
D. 10 examples covering every variation the codebase contains.

27

Which two outcomes do well-chosen few-shot examples produce? (Select two.)

A. Greater consistency across similar inputs.
B. Better handling of ambiguous cases.
C. Elimination of semantic errors in structured output.
D. Guaranteed schema validity.

28

The review agent must emit findings as structured JSON for the PR-comment bot. What mechanism does the exam expect?

A. A forced tool call with a JSON schema.
B. A prompt instruction to reply in JSON only.
C. A regex post-processor over the model's prose.
D. A markdown table the bot parses.

29

All findings are schema-valid, but some cite line numbers that don't exist in the diff. What is needed?

A. A validation step checking each line_number against the diff, discarding or re-querying invalid findings.
B. A stricter integer constraint on line_number.
C. A larger max_tokens.
D. tool_choice: "any".

30

The team wants a /review-pr workflow every developer can run locally too. What do they create?

A. .claude/commands/review-pr.md, committed.
B. ~/.claude/commands/review-pr.md
C. An entry in a commands array in .claude/config.json.
D. A --review-pr CLI flag.

Scenario C — Customer Support Resolution Agent (second visit)

Same system as Mock 1 Scenario A: an Agent SDK support agent with get_customer, lookup_order, process_refund, and escalate_to_human. This time the focus is the agentic loop itself, delegation, and reliability under load.

31

Which condition correctly terminates the agentic loop?

A. stop_reason is not "tool_use".
B. The response contains no tool_use blocks and the text is non-empty.
C. A maximum iteration count is reached.
D. The response text contains a completion phrase.

32

After receiving tool results, what must be appended to messages and in what form? (Select two.)

A. The assistant's response content, verbatim.
B. A single user message containing every tool_result block, each keyed to its tool_use_id.
C. A summary of the assistant's response, to save context.
D. One user message per tool result, in call order.

33

Which are loop anti-patterns? (Select three.)

A. Using an arbitrary iteration cap as the primary stop condition.
B. Terminating when the assistant response contains any text.
C. Terminating after the first tool result is returned.
D. Appending the assistant's response content verbatim before the tool results.

34

The team is deciding between a model-driven agent and a pre-configured decision tree for triage. Which favours the model-driven approach?

A. The set of enquiry types is open-ended and evolves.
B. Every path must be auditable in advance.
C. There are exactly four enquiry types, all known.
D. Regulators require a documented, fixed flow.

35

The team adds a specialized returns subagent. How should the main agent delegate to it?

A. Via the Task tool, with "Task" present in allowedTools.
B. By calling the subagent's tools directly.
C. By setting delegate: true in the configuration.
D. By starting a new session for the subagent.

36

The returns subagent's AgentDefinition needs three things filled in. Match them. (Select three.)

A. description — used by the coordinator to decide when to route here.
B. prompt — the subagent's own instructions.
C. Scoped tools — only what this role requires.
D. parallel — whether it may run concurrently.

37

A delegating prompt reads "Handle the return for the order we discussed." The subagent asks which order. Why?

A. It lacks lookup_order.
B. Subagents have isolated context and inherit nothing; the referent is undefined.
C. The instruction is lost in the middle of the prompt.
D. Its model is too small.

38

Two enquiries in one conversation are fully independent: a refund on order A and a shipping question on order B. How does the agent handle both concurrently via subagents?

A. Two Task calls in a single assistant response.
B. parallel: true in the configuration.
C. One Task call per turn.
D. A single Task call listing both.

39

Which delegating prompt style does the exam prefer?

A. State the goal and the quality criteria for a successful result.
B. Enumerate the procedural steps the subagent must follow in order.
C. Provide the coordinator's full conversation history.
D. Provide only the customer's raw message.

40

An escalation handoff must be built. What goes in it?

A. Customer ID, root cause, refund amount, recommended action.
B. The full transcript.
C. A one-line reason.
D. The transcript plus per-action confidence scores.

41

Under load the team wants to cap runaway tool use without breaking correct termination. What is the right relationship between the cap and the loop condition?

A. stop_reason remains the primary condition; the cap is a safety backstop that logs and escalates when hit.
B. The cap becomes the primary condition, checked before stop_reason.
C. Remove the cap; stop_reason is sufficient.
D. Cap at one tool call per turn.

42

An agent handling a 30-turn conversation is producing responses that contradict commitments made in turn 3. Which two changes are appropriate? (Select two.)

A. Maintain a persistent case-facts block outside the summarized history.
B. Trim verbose tool outputs so real content isn't crowded out.
C. Switch to a larger context window.
D. Drop the oldest turns once a threshold is reached.

43

A refund tool call fails with a transient network error. What should happen?

A. Return isError: true, errorCategory: "transient", isRetryable: true; the agent retries.
B. Return isError: true, isRetryable: false; escalate.
C. Return isError: false with an empty result.
D. Terminate the conversation.

44

The team wants to be certain a refund is never issued twice for one order. Which mechanism?

A. A PreToolUse gate checking a persisted record of completed refunds for that order.
B. A prompt instruction not to issue duplicate refunds.
C. A PostToolUse hook logging refunds for reconciliation.
D. A few-shot example showing a duplicate request being declined.

45

A developer returns to a support-agent debugging session after the weekend. The recorded reasoning is valuable but the tool results in it are stale. Best action?

A. Start a fresh session with a structured summary of the prior conclusions.
B. --resume the prior session.
C. fork_session from the prior session.
D. /compact and continue.

Scenario D — Multi-Agent Research System (second visit)

Same system as Mock 1 Scenario C: a coordinator delegating to web-search, document-analysis, synthesis, and report-generation subagents. This time the focus is orchestration topology, decomposition quality, and output fidelity.

46

Which statement describes the required communication topology?

A. Hub-and-spoke: all inter-subagent communication routes through the coordinator.
B. Mesh: subagents communicate directly for efficiency.
C. Chain: each subagent passes results to the next.
D. Broadcast: every subagent sees every other's output.

47

Which two statements about subagent context are correct? (Select two.)

A. Subagents have isolated context.
B. Subagents inherit nothing from the coordinator's conversation.
C. Subagents inherit the coordinator's system prompt and history.
D. Subagents share a common context pool.

48

The coordinator decomposed "assess this company's competitive position" into eight subtasks, each covering a single product feature. The report is detailed but misses pricing, distribution, and regulation. Root cause?

A. Overly narrow decomposition.
B. Insufficient subagent count.
C. A synthesis failure.
D. Inadequate tool access.

49

Which decomposition pattern fits a question whose required sub-areas can only be determined after an initial survey?

A. Dynamic, adaptive decomposition — survey first, then decide the subtasks.
B. Prompt chaining with a fixed sequence of stages.
C. A pre-configured decision tree over question types.
D. Fixed partitioning into five standard business dimensions.

50

Which pattern re-delegates after a coverage check reveals a gap?

A. Iterative refinement.
B. Scope partitioning.
C. Dynamic subagent selection.
D. Prompt chaining.

51

The report-generation agent produces a single 40-page document. Reviewers say they cannot find the answer to the client's actual question. What should change?

A. Lead with key findings, then use explicit section headers for the rest.
B. Shorten the report to 10 pages.
C. Add a table of contents.
D. Increase the report agent's context window.

52

A synthesis pass must combine financial comparisons, a regulatory timeline, and a list of technical risks. How should each render?

A. Financials as tables, the timeline as prose, technical risks as lists.
B. All three as prose, for readability.
C. All three as tables, for consistency.
D. All three as lists, for scannability.

53

A subagent returns four paragraphs of reasoning before its findings. Over a long run the coordinator's output becomes vague. Which two changes help most? (Select two.)

A. Specify a structured return format — findings only, as {claim, source, date} records.
B. Have the coordinator maintain a scratchpad file of accumulated findings.
C. Increase the coordinator's context window.
D. Have subagents report directly to the synthesis agent.

54

Reviewing a 60-file codebase for a client's technical due diligence, a single reviewing agent misses issues in files it read early on. Best approach?

A. Per-file passes, then a separate cross-file integration pass.
B. A larger context window so all 60 files fit.
C. Reading the files in dependency order.
D. Extended thinking on a single pass.

55

The final report must disclose what was not covered. Where does that information come from?

A. Coverage annotations propagated from each subagent's structured returns, including access failures and gaps.
B. The synthesis agent inferring gaps from what's present.
C. A final review pass comparing the report against the original question.
D. The client noticing and asking.

56

Two subagents were asked the same question and returned conflicting figures for market size. What should the synthesis agent do?

A. Report both with attribution, publication dates, and any methodological difference.
B. Take the figure from the more capable subagent.
C. Use the more recent figure.
D. Average them.

57

The coordinator must decide which subagents to spawn for a question about a heavily regulated industry, where a regulatory specialist matters and a distribution specialist doesn't. Which pattern?

A. Dynamic subagent selection.
B. Scope partitioning.
C. Iterative refinement.
D. Prompt chaining.

58

A subagent's prompt instructs it to "research thoroughly and return whatever seems relevant." Findings are inconsistent between runs. Most effective fix?

A. State the goal and explicit quality criteria for a complete result, plus the required return format.
B. Instruct it to be more thorough.
C. Lower its temperature.
D. Run it three times and merge.

59

Which behaviour is correct when one of four document collections is inaccessible?

A. Return the three successful result sets plus a structured error for the fourth, with category, what was attempted, and alternatives.
B. Return a generic "search unavailable" error.
C. Return only the three successful result sets.
D. Terminate the workflow.

60

A multi-week research engagement spans several sessions. What must the state export carry to keep later work consistent with earlier work?

A. The conventions and decisions established earlier, not just the list of completed items.
B. A timestamp per completed item.
C. The full conversation history of each prior session.
D. A confidence score per completed item.

End of Mock Exam 2. Record your answers before checking the key.