Claude Certified Architect – Foundations Practice Question 42: Subagent Structured Status
Claude Certified Architect, Foundations · Test 5 · question 2 of 10
A coordinator agent fans out four research subagents, one per data source, and merges their findings into a report. In production, the report sometimes omits a whole section with no indication of why. Investigation shows that when a source API times out, the subagent catches the error and returns an empty string as its final message, which the coordinator treats as "nothing relevant found." The team wants the coordinator to retry transient failures, and to state clearly in the report which sources could not be covered. What is the best fix?
Show answer and explanation
Answer:
The root cause is that a failure and a legitimate "no results" outcome look identical to the coordinator. A subagent's only channel back is its final message, so the contract for that message must distinguish them. A structured result with status, error type, retryability and partial findings lets the coordinator apply deterministic logic: retry transient failures with backoff, keep partial findings, and disclose gaps in the report. This matches Anthropic's multi-agent guidance that errors compound in long-running agent systems, that agents should be told when a tool is failing so they can adapt, and that retry logic should be combined with deterministic safeguards.
A is wrong because a turn limit does not fix an ambiguous signal; the coordinator would loop on a source that is genuinely down and still could not tell "empty" from "failed". B is wrong because aborting the whole run discards the three sources that worked, and it makes every transient timeout user-visible. C is wrong because re-running everything wastes cost and still cannot tell a failure from an empty result; "longer" is not a quality signal and a failed run would simply keep returning nothing.
Reference: How we built our multi-agent research system