Appendix C. Exam day guide
Time allocation
Forty questions in 70 minutes gives one minute 45 seconds per question. Analytical questions have longer passages and take more time, while concept questions are quicker. Allocate time as follows.
- Attempt every question once during the first 50 minutes. If you are not confident within one minute 30 seconds, mark it and move on.
- Spend the next 15 minutes returning to marked questions.
- Use the last five minutes to check your answer sheet. Leave no blanks and verify that changed answers are marked correctly.
Solving analytical questions
- Read the question first. Check what you must select: the most appropriate option, the first action, or two options, for example.
- List the bold conditions. Constraints determining the answer are usually bold. Note cost, latency, reversibility, trust boundaries and operating conditions such as retry availability separately.
- Find the constraint each option violates. Wrong answers usually satisfy one constraint but violate another. Eliminate an option once you find even one violation.
- Ask, “What if this condition changed?” Consider whether a remaining option becomes wrong when one condition changes. This confirms the link between the condition and the answer.
Common wrong-answer patterns
| Pattern | Appearance | Relevant chapter |
|---|---|---|
| Over-engineering | Autonomous or multi-agent systems for fixed flows | Chapter 1 |
| Mistaking agreement for verification | Relying on agreement among several agents | Chapter 1 |
| Blaming tool counts alone | Splitting agents to solve tool similarity | Chapters 1 and 2 |
| Mistaking hints for guarantees | Basing permission decisions on tool annotations or descriptions | Chapter 2 |
| Token passthrough and broad scopes | Forwarding received tokens or granting all permissions initially | Chapter 2 |
| Treating window size as a universal solution | Covering management problems with a larger context window | Chapter 3 |
| Relying on memory | Entrusting resumption or progress state to model memory or summaries | Chapter 3 |
| Confusing statements with state | Treating the agent's completion claim as success | Chapter 4 |
| Rigid grading | Exact trajectory matching failing other correct paths | Chapter 4 |
| Misused metrics | pass@k without retries; averages masking critical regressions | Chapter 4 |
| Mistaking advice for enforcement | Believing prompt instructions block dangerous actions | Chapter 5 |
| Excessive or absent approval | Approval for every action, or only keeping logs | Chapter 5 |
| Treating guardrails as universal protection | Believing concurrent input guardrails prevent tool side effects | Chapter 5 |
| Unconditional retries | Retrying unchanged business errors or failures with unknown execution status | Chapters 2 and 5 |
Tips by question type
- Single choice: If two candidates remain, apply each bold condition. Options using an unnecessarily complex tool or reaching conclusions without measurement are usually wrong.
- Multiple choice: Judge each item in the
<Statements>box independently, then select the combination containing the stated number of correct items. Reconsider judgments if the count does not match. - Statement judgment: Judge conclusion and reason separately. If the grounds after “because” contradict the documentation, the statement is incorrect even if its conclusion is right.
- Ordering: Check whether each step needs the preceding step's output. Recovery before containment or scoping, and re-execution before checking state, do not work.
- Matching: Eliminate certain pairs first. Remember one item on the right remains, and verify it matches nothing on the left.
Key point
When stuck, return to three questions: “Is this the simplest structure? Is enforcement on the execution path? Is the judgment based on measured records?” The option passing all three is usually correct.