Mock exam 1: Architect – Foundations
A timed practice exam in the official style: scenario-framed, domains hidden, answers at the end. Unlike the practice questions, this measures readiness rather than teaching. Original questions written against the public blueprint and the six published scenarios; not items from the live exam, which is covered by a non-disclosure agreement.
How to take it. 15 questions, 30 minutes, closed book: no notes, no documentation, no Claude. The real exam presents 4 scenarios drawn from the published 6; this mock samples across all of them. Score yourself with the key and the readiness table afterwards.
Scenario A: customer support resolution agent
An Agent SDK agent handles returns, billing disputes, and account issues through MCP tools: get_customer, lookup_order, process_refund, escalate_to_human.
1. The agent occasionally answers billing questions using an order it looked up three turns earlier for a different customer. What is the architectural fix?
- A. Increase the context window
- B. Scope each conversation's context to one customer and clear or isolate prior lookups when the subject changes
- C. Remove the lookup_order tool
- D. Instruct the model to be careful about customer identity
2. process_refund fails intermittently because the billing backend times out. What should the tool return so the agent can behave sensibly?
- A. An empty string
- C. A structured error with a category and a retryable flag, so the agent can retry or escalate deliberately
- B. A message asking the user to try later
- D. The exception stack trace
3. Which condition should route a ticket to escalate_to_human rather than continue autonomous resolution?
- A. The customer writes in a language the agent supports
- D. The request falls outside documented policy, or the agent cannot make progress after a bounded number of attempts
- B. The order is more than thirty days old
- C. The conversation exceeds five turns
4. The agent sometimes calls get_customer when it should call lookup_order. What is the most effective change?
- B. Reorder the tools in the configuration
- A. Rewrite both tool descriptions so each states precisely when it applies and how it differs from the other
- C. Merge them into one tool with a mode flag
- D. Increase the model's temperature
Scenario B: Claude Code for a development team
A team uses Claude Code across a monorepo for generation, refactoring, and review.
5. Backend conventions should apply only when backend files are touched. Where do they belong?
- A. In the root CLAUDE.md so they are always available
- B. In a path-scoped rule under .claude/rules/ with frontmatter matching backend paths
- C. In each engineer's user-level configuration
- D. In the repository wiki
6. When is plan mode the appropriate choice over direct execution?
- A. For every change, since planning is always safer
- C. When the change is large or risky enough that a human should review the approach before edits are made
- B. Only when the model is unfamiliar with the language
- D. When the context window is nearly full
7. A repeated multi-step review procedure should be available to the whole team as a single invocation. What is the right mechanism?
- A. A shared document describing the steps
- D. A custom slash command or skill committed to the repository
- B. A longer CLAUDE.md
- C. A saved chat transcript
Scenario C: multi-agent research system
A coordinator delegates to search, analysis, synthesis, and report subagents that produce cited reports.
8. Subagents return full source documents to the coordinator, which then exceeds its context limit. What is the correct design change?
- B. Give the coordinator a larger model
- A. Have subagents return structured findings with citations while raw sources stay at the edge or in a scratchpad
- C. Reduce the number of subagents to one
- D. Truncate each document to its first page
9. Final reports attribute claims to the wrong sources even though each subagent is individually accurate. What is the likely cause?
- A. The synthesis model is too small
- B. The handoff format does not carry claim and source together, so provenance is lost between stages
- C. The search subagent returns too many results
- D. Reports are generated before analysis completes
10. A subagent fails permanently on one source. What should the coordinator receive?
- A. Nothing, so the report proceeds
- C. A propagated error with enough context to decide between retry, exclusion with a noted gap, or abort
- B. A silent retry loop until it succeeds
- D. A generic failure flag with no detail
Scenario D: Claude Code in CI/CD
Claude Code runs automated review on pull requests and posts feedback.
11. The reviewer must emit results a pipeline can parse. Which invocation is correct?
- A. Interactive session with output redirected to a log
- D. Non-interactive print mode with JSON output and a schema
- B. Plan mode with an approval step
- C. A cron job that emails the transcript
12. Reviews flag many low-value style nits and engineers now ignore them. Which change addresses this?
- B. Post findings to a separate channel
- A. Define explicit review criteria and a severity bar in the prompt, with examples of in-scope and out-of-scope findings
- C. Run the review only on large pull requests
- D. Cap output at five findings
Scenario E: structured data extraction
Documents are parsed into validated records for downstream systems.
13. Some invoices legitimately lack a delivery date. How should the schema express this?
- A. Make the field required and let extraction fail
- B. Make the field nullable so absence is recorded explicitly, with conditional validation where the document type requires it
- C. Omit the field from the schema
- D. Default it to today's date
14. Which loop gives the most reliable extraction output?
- A. Generate, then manually review every record
- C. Enforce a JSON schema through tool use, validate, and on failure retry with the validation error provided to the model
- B. Generate three times and take the majority
- D. Generate with temperature zero and accept the result
15. Your team reports 97% accuracy measured only on records that passed validation. Why is that number misleading?
- A. Accuracy should be measured per field, not per record
- D. It excludes the failures, so it measures where the system already succeeded rather than the true rate across all documents
- B. 97% is too low to report
- C. Validation and accuracy are unrelated
Answer key
| # | Answer | Domain | Why |
|---|---|---|---|
| 1 | B | Context Management & Reliability | Cross-customer contamination is a context-scoping defect, not a window-size problem |
| 2 | C | Tool Design & MCP Integration | Structured errors with categories and retryable flags are what let an agent recover deliberately |
| 3 | D | Context Management & Reliability | Policy gaps and lack of progress are the documented escalation triggers; turn counts and order age are not |
| 4 | A | Tool Design & MCP Integration | Tool selection runs on descriptions, so differentiation is the fix for confusion between similar tools |
| 5 | B | Claude Code Configuration & Workflows | Path-scoped rules load conditionally, which is exactly the stated requirement |
| 6 | C | Claude Code Configuration & Workflows | Plan mode is a human review gate for approach, warranted by risk rather than universally |
| 7 | D | Claude Code Configuration & Workflows | Repeatable team procedures belong in committed commands or skills, not prose |
| 8 | A | Context Management & Reliability | Delegation exists to keep bulk at the edge and move distilled results |
| 9 | B | Context Management & Reliability | Provenance survives handoffs only when the format carries claim and source together |
| 10 | C | Context Management & Reliability | Errors must propagate with enough context for the coordinator to choose a recovery path |
| 11 | D | Claude Code Configuration & Workflows | CI requires non-interactive invocation with structured, schema-validated output |
| 12 | A | Prompt Engineering & Structured Output | Precision comes from explicit criteria and examples that draw the signal boundary |
| 13 | B | Prompt Engineering & Structured Output | Nullable fields record legitimate absence; required fields force fabrication or failure |
| 14 | C | Prompt Engineering & Structured Output | Schema enforcement plus a validation-retry loop is the reliability pattern the guide describes |
| 15 | D | Context Management & Reliability | Measuring only validated records is a biased sample; calibration needs coverage of all documents |
Interpreting your score
| Raw score | Reading |
|---|---|
| 13 to 15 | Comfortable. Review misses, reread the six scenarios, and book the exam |
| 10 to 12 | Close. Work the weak domains, then do the guide's own preparation exercises |
| 7 to 9 | More study needed. Build an Agent SDK agent and configure Claude Code on a real project before testing |
| 6 or fewer | Start from the exam guide, including its appendix of named technologies |
This scale is a study aid, not a predictor. The real exam reports a scaled score from 100 to 1,000 with 720 to pass.
Tally misses by domain; two or more in one domain marks your next study target, as the real score report's percent-correct breakdown would.
Written by the maintainer for self-assessment. Study guide · Study notes · Mock 2 · Mock 3 · Practice questions · Repository index