Domain 1 · Cumulative Retrieval Quiz (27%)

Domain 1 — Test Yourself

Eight scenario questions, interleaved across Tasks 1.1–1.7. No looking back first — retrieval is the point. Click an answer to grade + explain.

Aim: 7/8+. Below that, re-read the flagged lesson before Domain 2.

1. Your agentic loop occasionally terminates while Claude still intends to call a tool. The correct control condition continues the loop whenever stop_reason is:

"tool_use", and terminates only on "end_turn"
"end_turn", and terminates only on "tool_use"
absent, terminating once any assistant text is present
repeated, terminating at a fixed iteration ceiling

Correct: option 1. Continue on tool_use, stop on end_turn. (L1)

2. A coordinator must be able to spawn subagents. This requires that its configuration:

Sets tool_choice to the value "any" on every request
Includes "Task" in its allowedTools list
Shares its full history with each subagent by default
Runs all subagents within one single shared session

Correct: option 2. The Task tool spawns subagents; allowedTools must include it. Subagents are context-isolated (option 3 is false). (L2)

3. Refunds above $500 must never be issued autonomously. The reliable enforcement mechanism is a:

System prompt clause stating the $500 limit is a hard rule
Few-shot set showing refusals of over-limit refund requests
Tool-call interception hook that blocks and redirects over-limit
Escalation instruction appended to the agent's response format

Correct: option 3. A guarantee needs deterministic code (interception hook). Prompt/few-shot are probabilistic. (L3)

4. A research report has coverage gaps although every subagent succeeded and the coordinator's log shows a narrow topic split. Fix the:

Relevance filter inside the document analysis subagent
Query breadth used by the web search subagent alone
Gap-detection instructions given to the synthesis agent
Coordinator's task decomposition that assigned narrow scope

Correct: option 4. Subagents executed their given scope correctly; the upstream decomposition is the root cause. (L2)

5. You want two subagents to run at the same time to cut latency. You should have the coordinator:

Emit multiple Task calls within a single response
Emit one Task call per turn across separate responses
Set a high max_tokens so both fit in one output
Fork the session once for each subagent it needs

Correct: option 1. Parallelism = multiple Task calls in one response; separate turns run sequentially. (L2)

6. A 14-file pull request review gives inconsistent depth and contradictory findings in one pass. The best restructure is to:

Run per-file local passes, then a cross-file integration pass
Move to a larger context window so all files fit in one pass
Require developers to submit no more than four files per PR
Run three full passes and keep issues found in at least two

Correct: option 1. The cause is attention dilution; split into focused passes. A bigger window does not fix attention quality; the others shift burden or suppress real bugs. (L4)

7. Escalating to a human agent who cannot see the transcript, you should hand off:

A note that the customer is frustrated and wants help
A structured summary: ID, root cause, amount, recommended action
The raw conversation log for the human to read fully
Only the final unresolved question the customer asked

Correct: option 2. Structured handoff lets the human act without re-interviewing the customer. (L3)

8. You resume work after files you analyzed were heavily rewritten. Most reliable path:

Resume the named session to retain all prior tool results
Fork the session to branch from its cached analysis baseline
Start fresh with a structured summary; prior reads are stale
Resume and rely on the model to notice the changes itself

Correct: option 3. Stale tool results make fresh+summary safer than resume/fork, which both inherit outdated reads. (L4)

Scoring 8/8 — move to Domain 2. 6–7 — review the lesson(s) behind any miss. ≤5 — re-read Domain 1 lessons before continuing; this is the heaviest domain (27%).

Tell your teacher your score and which numbers you missed. I'll log it to your learning record, target those weak spots, and then build Domain 2.