CCAR-P - Solution Design & Architecture (17% of the exam) - Section 1.4

Design multi-agent systems and orchestration strategies.

Coordinating several agents: an orchestrator that delegates to specialised subagents, how context passes between them, how results are aggregated, and how failures in one agent are contained. Candidates should judge when parallel specialised agents earn their coordination cost and when a single agent is enough.

orchestrator and subagentscontext passing between agentsresult aggregationfailure containmentcoordination cost

Practice question for this objective

Free sampleSolution Design & Architecturehard

A state benefits agency runs an orchestrator that sends each application to three eligibility subagents in parallel, one each for housing, childcare and disability support, then merges their findings into one determination letter. During a four-day outage of the housing register, the housing subagent could not complete its lookups, yet no run was marked as failed and no letter mentioned a problem. Appeals later showed that 11 percent of letters issued in that window wrongly stated the applicant was ineligible for housing support, while the childcare and disability findings in the same letters were correct. Which cause does this evidence point to most strongly?

AThe merge step treats a branch that returned no finding as a negative finding, so a failed lookup became a refusal Correct
BThe orchestrator briefed the housing subagent with too little applicant detail to reach a sound eligibility decision
CThe parallel subagents interfered with each other, so another branch's output overwrote the housing subagent's result
DThe housing subagent's model misread the eligibility rules, so it reached refusals the policy text does not support

Aggregation must distinguish a failed branch from a negative result, or a subagent's failure is silently converted into a wrong final decision. In a fan-out design, failure containment depends on the aggregator knowing which branches succeeded. The clue is the combination of an upstream outage, zero failed runs and confidently wrong refusals confined to one branch. That pattern means the housing branch's failure was passed upward as an empty or absent finding and merged as if it were a decision. The structural fix is for branches to return an explicit failure status and for the merge step to hold or flag the letter rather than issue a determination.

Why A is correct: No run failed and no letter flagged a problem, yet the outage produced refusals, which means the failure was absorbed somewhere and reinterpreted. Aggregation that cannot tell a failed branch from a negative result converts an outage into a wrong determination instead of containing it.

Why B is wrong: Thin briefs are a common multi-agent defect, but the briefing did not change during the outage and the wrong refusals are confined to that window. A briefing problem would show up before and after the outage as well.

Why C is wrong: Shared mutable state between parallel branches can corrupt results, but the childcare and disability findings were correct and the errors line up with the register outage, not with concurrency. Cross-branch interference would not track an upstream outage.

Why D is wrong: Blaming model reasoning is tempting when decisions are wrong, but the errors began and ended with the outage while the model and rules were unchanged. The subagent could not complete its lookups at all, so it had nothing to misread.

See more CCAR-P practice questions, answers explained.

Exam traps in Solution Design & Architecture

Answers that look right on this material and are not. Each one is a distractor from a different question in the CCAR-P bank for this domain.

Three voters is too small a panel, and widening it to five reviewers would push unanimous errors well below one reviewer's rate.
Why it is wrong: Tempting because majority voting does cut error as the panel grows when voters err independently. It is wrong here because every reviewer reads the same incomplete excerpt with the same prompt and model, so two more copies would miss the same exclusion criteria and add cost without adding independent evidence.
Cap each subagent's return at its first 200 log lines so that all three outputs fit comfortably in the coordinator's context
Why it is wrong: This is tempting because it directly reduces context pressure. It is wrong because the first lines retrieved are not necessarily the relevant ones, so the cap discards evidence the diagnosis may depend on and breaks the traceability requirement for anything beyond the cut.
The summary subagent's model tier is too small to carry a contraindication reliably into the final summary text
Why it is wrong: A capability gap is a tempting explanation whenever output quality drops, but the model did not change and the omissions began exactly when the handoff format changed. Upgrading the tier would not restore a flag that never reaches the summary subagent.

Examworthy is not affiliated with or endorsed by Anthropic. Original, blueprint-aligned practice material only.