Vault / wiki/301/practice/domain-1-agents.md
updated 2026-05-28Practice — Domain 1: Agentic Architecture & Orchestration (27%)
18 scenario-based MCQs. Answer key + explanations at the bottom.
Q1
You're building a system that translates marketing copy into 12 languages and posts each to a different regional CMS. The translation prompt, posting target, and validation rules are fixed at design time. Which architecture is most appropriate?
A. Orchestrator–workers with one subagent per language B. Prompt chaining (translate → validate → post) per language C. A single autonomous agent with all 12 CMS tools available D. Evaluator–optimizer between translator and editor
Q2
A customer-support bot must decide whether each incoming message is a refund, billing question, technical issue, or sales inquiry, then route it to a handler with the right tools. What pattern is this?
A. Orchestrator–workers B. Routing C. Evaluator–optimizer D. Prompt chaining with gates
Q3
You need a research agent that produces a 5-page report from ~30 web sources. The structure of the report depends on what the sources reveal — you can't predetermine the sections. Which pattern fits best?
A. Routing with one handler per source type B. Prompt chaining: outline → draft → polish C. Orchestrator–workers with research/analyze/write subagents D. Voting: run the report 5 times and merge
Q4
A financial team wants Claude to write 10K product descriptions overnight, each from a structured input row. There's no interactive requirement; cost is the constraint. Which combination is most aligned with the certification's expected best practice?
A. Streaming messages.create calls with Sonnet
B. The Message Batches API with a forced-tool extraction schema on Haiku
C. A single long agent loop iterating row-by-row
D. Orchestrator–workers with one subagent per row
Q5
A subagent in an orchestrator–workers pattern needs the orchestrator's earlier reasoning to do its job correctly. What's the right move?
A. Share the orchestrator's full message history with the subagent B. Include only the orchestrator-prepared brief (a self-contained prompt with the necessary facts) C. Have the subagent call the orchestrator back when it needs context D. Store the orchestrator's state in a shared mutable variable accessible to all subagents
Q6
You're designing an evaluator–optimizer for code reviews where the evaluator flags bugs and the optimizer revises. Which condition is most important for this pattern to work well?
A. The evaluator must run on a larger model than the optimizer B. There must be clear evaluation criteria the evaluator can apply consistently C. The two LLMs must use different tool sets D. The pattern requires plan mode
Q7
A team is concerned their classifier-and-handler system is over-engineered. The single Sonnet prompt handles all 4 customer service domains with conditional rules. Symptoms: the prompt is 2,500 tokens, accuracy on the rarest domain has dropped over time, and adding a 5th domain made things worse across the board. What's the most likely root cause?
A. The model size is too small B. Dilution from too many conditional rules in one prompt C. Missing few-shot examples D. Lack of structured output
Q8
"Subagents must not share mutable state directly." Why is this the convention?
A. The Agent SDK doesn't support shared variables B. It creates races and contradictions; results must flow through the orchestrator for integration C. Each subagent runs on a different model and can't read another's outputs D. Subagents are stateless by definition
Q9
An autonomous Claude Code agent has been looping for 47 iterations on a refactor. It keeps editing and re-testing without converging. What reliability primitive is missing?
A. Prompt caching B. A capped iteration count with a fallback to escalate or stop C. Extended thinking D. More tools
Q10
You're tasked with a content moderation system that must minimize both false positives and false negatives for sensitive material. Which agentic pattern is best aligned?
A. Routing — one handler per content type B. Prompt chaining — moderate, then second-pass classify C. Parallelization (voting) — run multiple classifiers and combine D. Evaluator–optimizer — moderator generates verdict, evaluator critiques
Q11
Anthropic's "Building Effective Agents" recommends investing in tool documentation roughly equal to investment in prompts. The SWE-bench team famously switched from relative to absolute file paths to fix a class of errors. What general principle does this illustrate?
A. Larger context windows beat better prompts B. Agent-computer interface (ACI) design is as important as prompt engineering C. Filesystem tools should always require absolute paths D. Tools should never accept ambiguous inputs
Q12
A research orchestrator has spawned three workers in parallel: search, analyze, write. The write worker has started before analyze finishes, producing incoherent output. What's wrong with the design?
A. The workers should be sequential because write depends on analyze's output B. The workers should all be on the same model C. The orchestrator needs more tools D. The pattern should be evaluator-optimizer
Q13
For a tax-preparation agent that walks a user through filing a return, the conversation is long and the state (income, dependents, deductions) changes over many turns. Which context-management approach in an agent loop best avoids ballooning costs while preserving state?
A. Replay full conversation history every turn — caching will handle the cost B. Maintain a compact JSON state object in the system prompt and prune old turns C. Switch to a smaller model after 10 turns D. Use sliding window only (last 5 turns)
Q14
When does "agent" become the right choice over "workflow"?
A. Whenever the LLM uses any tool B. Whenever the task has more than three steps C. When step counts cannot be predicted and the system needs to decide next actions based on environment feedback D. Whenever the user asks for an "AI agent"
Q15
Which item below is NOT one of the three core principles Anthropic articulates for building agents?
A. Simplicity in design B. Transparency of planning steps C. Tool documentation and testing D. Always use the largest model available
Q16
A multi-agent system uses Opus for the orchestrator and Haiku for the workers. Why is this typical?
A. Opus is required for tool use; Haiku is not B. Orchestrators need stronger reasoning; workers are scoped and benefit more from speed and cost C. Haiku can't run in subagents D. Opus must run as a subagent for security
Q17
You're shipping a chat support assistant that occasionally needs to call an external process_refund tool. The model has chosen to call it. What's the right confirmation pattern?
A. Auto-confirm if the refund is under $50, ask the user otherwise B. Always require explicit user confirmation from the host application before executing the side-effecting call C. Have a second LLM verify the refund details before calling D. Cache the refund decision and let the user undo via support
Q18
The orchestrator emits a plan and asks the user to approve before any worker runs. The user can edit the plan, then approve. Which Claude Code feature is this most aligned with?
A. Slash commands B. Hooks C. Plan mode D. Skills
Answers
Q1: B. Steps are fixed at design time → workflow, specifically prompt chaining (translate → validate → post). Orchestrator–workers (A) is overkill when the plan is static. Single autonomous agent (C) is the most overkill option. Evaluator–optimizer (D) doesn't match — there's no iterative quality loop required.
Q2: B. Routing is the canonical pattern for "classify, then dispatch to specialized handler." Orchestrator–workers (A) implies dynamic decomposition; this is a fixed lookup. Chaining (D) is sequential, not branching.
Q3: C. When the plan depends on the data and decomposition is dynamic, orchestrator–workers is the canonical pattern. (A) is wrong because the structure isn't predetermined. (B) imposes a fixed structure. (D) wastes resources on consensus when each pass would have different findings.
Q4: B. Message Batches API gives 50% off and is async-suitable (no SLA pressure). Forced tool use is the most reliable way to get structured output. Haiku is the cost-aligned model for bulk classification/extraction. (A) loses the batch discount. (C) is monolithic and slow. (D) wastes coordination overhead.
Q5: B. Subagents must be briefed with self-contained prompts. The parent's history is noise; only the relevant facts should be included. Sharing full history (A) destroys the context-isolation benefit. Callbacks (C) defeat the purpose. Shared mutable state (D) is explicitly an anti-pattern.
Q6: B. The evaluator-optimizer pattern works only when there are clear evaluation criteria; otherwise the evaluator's feedback is noise that drives the optimizer in random directions. Quote from the source: "Most effective when we have clear evaluation criteria." (A) is sometimes true but not the most important factor.
Q7: B. Classic dilution. Adding conditionals lowers Claude's weighting of any single rule. The fix is to route to specialized prompts. Bigger model (A) is a band-aid; the root cause is architectural.
Q8: B. Shared mutable state causes races and contradictions between workers. The discipline is that results flow up to the orchestrator, which integrates. (A) is incorrect — the issue isn't tooling. (C) isn't true. (D) is also not the design rationale.
Q9: B. Capped iteration counts are a basic agent-reliability primitive. The agent should also have a defined "fail loudly / escalate" path when the cap is hit. Tools (D) don't help if the loop logic is broken.
Q10: C. Voting (parallelization, same prompt N times, consensus) is the canonical pattern when both false-positive and false-negative rates matter — multiple independent passes reduce both error directions. Source: Anthropic's "Building Effective Agents" cites this exact example.
Q11: B. The principle is agent-computer interface (ACI) design quality matters as much as prompt quality. (C) is too specific; the general principle is broader. (A) is unrelated.
Q12: A. Parallelization requires independent subtasks. Write depends on analyze's output, so they must be sequential. The parent should plan the dependency.
Q13: B. State-object pattern is the canonical compaction strategy for transactional agents. (A) is wasteful even with caching because content changes invalidate the cache. (D) loses critical state from earlier in the flow.
Q14: C. From the source: agents handle open-ended problems where step counts cannot be predicted, and require ground truth from the environment at each step. (A) confuses tool use with autonomy. (B) is arbitrary.
Q15: D. The three principles are simplicity, transparency, and tool documentation and testing. "Always use the largest model" contradicts simplicity and is not a principle.
Q16: B. Orchestrators reason about plans; workers execute focused tasks. Haiku is cheap and fast — ideal for many parallel narrow tasks. Opus's reasoning is wasted on a tightly-scoped subagent. (A), (C), (D) are factually incorrect.
Q17: B. Destructive / side-effecting tools require host-level user confirmation before execution. Auto-confirm under a threshold (A) violates least-privilege and audit principles. Second-LLM verification (C) doesn't replace human consent for a side-effecting action.
Q18: C. Plan mode is the Claude Code feature where the agent proposes a plan, the user reviews/approves, then execution proceeds. (A)/(B)/(D) are different primitives.