CCAR-F - Agentic Architecture & Orchestration (27% of the exam) - Section 1.6
Design task decomposition strategies for complex workflows.
Fixed sequential pipelines (prompt chaining) suit predictable multi-aspect reviews; dynamic adaptive decomposition suits open-ended investigation where subtasks emerge from intermediate findings. Candidates should split a large code review into per-file local analysis passes plus a separate cross-file integration pass to avoid attention dilution, and decompose an open-ended task by first mapping structure, identifying high-impact areas, then building a plan that adapts as dependencies surface.
prompt chainingdynamic adaptive decompositionper-file versus cross-file passesattention dilutionadaptive investigation plans
The Atlas Research coordinator writes its complete subtask list for a research question before any subagent runs, then dispatches that list stage by stage without revising it. On questions whose shape is known in advance the reports score well. On 45 open questions where the first web search turns up an unexpected regulatory ruling, the finished report covers the originally planned lines of enquiry thoroughly and investigates the ruling nowhere, even though every subagent completed successfully and the coordinator's running summary mentions the ruling in passing. What best explains this?
AThe subtask list is dispatched stage by stage, so the ruling arrived after its own stage had already been consumed, and dispatching all remaining subtasks in a single batch instead would let the later stages see it.
BA decomposition fixed before any evidence exists can only encode the lines of enquiry that were knowable at planning time, so the coordinator has to re-plan against each stage's findings and mint subtasks the original plan could not have contained.check_circle Correct
CEach subagent returns its own findings without a judgement about their significance, so requiring every subagent to rate the importance of what it found will cause the ruling to be picked up and pursued.
DThe planning step ran without plan mode, so the coordinator produced an execution list rather than a reviewed plan, and enabling plan mode for the planning call will make the resulting decomposition cover unanticipated findings.
Open-ended research needs adaptive decomposition, because a plan fixed before evidence arrives cannot contain the branches that evidence reveals. A static decomposition is a commitment made at the point of least information. It works when the shape of the answer is known in advance, because the planner can enumerate the subtasks completely. On an open question the value often sits in a line of enquiry that becomes visible only once the first results return, and no amount of thoroughness within the original branches reaches it. Adaptive decomposition closes that gap by making the plan a living artefact the coordinator revises after each stage, adding, dropping or reordering subtasks in light of what came back.
Why A is wrong: It correctly notices that timing matters, but batching the remaining subtasks commits to the plan sooner rather than later, which makes the same problem worse rather than better.
Why B is correct: Correct. Adaptive decomposition treats the plan as revisable state updated by results, which is the only way a branch nobody anticipated can enter the workflow.
Why C is wrong: Self-reported significance is superficially attractive as a routing signal, but a rating changes nothing when no mechanism exists to create a new subtask in response, and self-assessed importance is a weak proxy for research value.
Why D is wrong: Plan mode is a real feature and naming it lends the option credibility, but it constrains an agent from acting while it plans and cannot give a planner evidence that has not been gathered yet.
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.
The per-document brief asks for figures and claims rather than for disagreements, so adding a sentence instructing each subagent to flag any figure it believes conflicts with other sources will surface the contradictions.
errorWhy it is wrong: Tightening the brief is a reasonable first instinct, but a subagent holding one document has no other document against which to judge a conflict, so the instruction asks for something it cannot perform.
Shorten the fixed investigation to two steps, a grep and a read of the best matching file, so that narrow questions stop paying for a sequence most of them never needed in the first place.
errorWhy it is wrong: It fixes the cheap half and worsens the expensive half, because the cross component questions were already answering from too little evidence and a shorter fixed plan starves them further.
A subagent inherits the coordinator's conversation but not its tool results, so the record was available to it as a summary line while the raw get_customer payload had already been stripped out.
errorWhy it is wrong: Tempting because partial inheritance sounds like a sensible cost saving, but no part of the coordinator's conversation is carried across, so there is no summary line for the subagent to read.