CCAR-F - Agentic Architecture & Orchestration (27% of the exam) - Section 1.2
Orchestrate multi-agent systems with coordinator-subagent patterns.
Hub-and-spoke architecture where one coordinator owns task decomposition, delegation, result aggregation and error handling, and subagents run with isolated context that does not inherit the coordinator's history. Candidates should design coordinators that select subagents by query requirements rather than always running the full pipeline, partition scope to minimise duplication, and run iterative refinement until coverage is sufficient. Overly narrow decomposition is the classic coverage failure.
hub-and-spoke coordinatorsubagent context isolationtask decomposition and delegationiterative refinement loopsscope partitioning across subagents
A coordinator splits a contact covering a damaged item and a duplicate charge across two subagents, briefing each one to resolve the customer's problem and passing both of them the full contact transcript. Reviewers find that both subagents investigated the duplicate charge, and in 27 of 200 such contacts the two returned different recommended amounts, with the coordinator relaying whichever result arrived last. Which TWO changes most effectively address this? Select TWO.
AGive each subagent a brief naming the one issue it owns and the records it may read, so that the two scopes do not overlap at all.check_circle Correct
BLet the two subagents exchange messages during the contact so they can agree between themselves which of them takes the duplicate charge.
CHave the coordinator reconcile the returned recommendations against the order record and decide the single amount that goes to the customer itself.check_circle Correct
DDispatch one subagent, then dispatch the second with the first one's returned transcript appended so that it can see what has already been covered.
EInstruct both subagents to state at the top of their returned answer which of the customer's issues they chose to work on during the dispatch.
Partition subagent scope explicitly in each brief and reconcile returned results at the coordinator rather than accepting the last one received. Two spokes given the same open instruction and the same transcript will converge on the most salient issue, which is why both took the duplicate charge. Naming one issue and its records per brief makes the fan-out disjoint, and a reconciliation step at the hub gives the contradictory case a defined resolution grounded in the order record instead of in arrival order.
Why A is correct: Correct. Both subagents chose the same issue because nothing in either brief assigned one, and a named scope per spoke is what makes a fan-out a partition rather than a duplication.
Why B is wrong: Tempting because negotiation would settle the overlap, but spoke-to-spoke traffic breaks the hub-and-spoke property that keeps each context isolated and moves the decision away from the component accountable for it.
Why C is correct: Correct. Arrival order is not evidence, so the hub needs an explicit merge step that weighs each result against the authoritative record before anything reaches the customer.
Why D is wrong: Tempting because it does inform the second subagent, but it doubles contact latency and still leaves the scope undefined, so the second one may reinvestigate the charge it was shown.
Why E is wrong: Tempting because it makes the overlap visible in review, but reporting a choice after the fact neither prevents the duplicated work nor tells the coordinator which of two amounts to trust.
Exam traps in Agentic Architecture & Orchestration
Answers that look right on this material and are not. Each one is a distractor from a different question in the CCAR-F bank for this domain.
Wrap the tool call in a retry with exponential backoff so a transient repository problem is given two further attempts before the agent concludes that the file list cannot be obtained.
errorWhy it is wrong: Tempting because retries are the standard remedy when a data fetch returns nothing useful and cost little on a short list call. It is wrong because the tool already succeeded and the scope genuinely contains no source files, so every retry returns the same result and the misreading persists three times over.
Raise the coordinator's context window allocation so that it can hold all four subagent transcripts at once and reconcile the overlapping sources during aggregation.
errorWhy it is wrong: Tempting because the coordinator does struggle to reconcile four overlapping result sets, but a larger window changes only what fits after the retrieval has already been paid for, so the duplicate fetches and the uncovered sub-questions both remain.
Each subagent receives the other's findings as they are produced, so whichever finishes second can deduplicate its source list against the first one's summaries.
errorWhy it is wrong: Tempting because both calls sit in one coordinator response and appear to be part of one shared piece of work, but there is no channel between sibling subagents through which findings could stream.