CCDV-F - Agents and Workflows (14.7% of the exam) - Section 1.1
Apply the principles, patterns, and tradeoffs of agent and workflow architecture.
Published skill weight 4.5 percent. Covers the decision criteria for building a workflow rather than an agent, the structure of manager and supervisor hierarchies, and the role subagents play in improving task execution. Candidates should be able to read a stated requirement and say which of the two shapes fits it, and justify where a supervisor layer earns its cost.
workflow versus agent decision criteriamanager and supervisor hierarchiessubagents for task executionagent architecture tradeoffs
Practice question for this objective
lock_openFree sampleAgents and Workflowsmedium
A team ships an internal assistant that answers one question at a time by calling a single search tool and summarising what it finds. Quality reviews show no correctness problem, but the team proposes adding a supervisor agent that would split each question across three subagents and merge their answers. The assistant has a stated response time budget that the current build already meets with little margin. What should the team do?
AAdd the supervisor and the three subagents, because parallel subagents finish sooner overall and the merge step raises answer quality on every question the assistant receives.
BAdd the supervisor but have it run the three subagents one after another, so that the merge step always receives complete results and the ordering of the sources stays stable.
CKeep one agent but raise the number of search results it may retrieve per question, so the supervisor layer becomes unnecessary and answers cover more of the corpus.
DKeep the single agent as it stands, and revisit a supervisor only if a measured failure appears that a coordinator with isolated subagent contexts would actually address.check_circle Correct
Add a manager or supervisor layer only when the task earns it, not as a default upgrade to a working single agent. Coordinator and subagent hierarchies pay for themselves when work can be decomposed into genuinely independent parts or when each part would otherwise pollute a shared context. A single-question lookup has neither property, so the extra planning, dispatch and merge calls consume the remaining latency margin and return nothing measurable.
Why A is wrong: Tempting because parallel work usually shortens wall-clock time, but the merge step is serial and the supervisor adds its own planning and synthesis calls, so total response time grows rather than shrinks on a single-question task.
Why B is wrong: Tempting because sequential execution removes merge race conditions, but running three subagents in series is the slowest option available and it breaks the response time budget while solving a problem the team has not measured.
Why C is wrong: Tempting because more retrieved context often helps, but it pushes more tokens through the same tight latency budget and it answers a coverage question that the quality reviews have not raised.
Why D is correct: Correct: a manager layer is justified by dynamic decomposition or context isolation needs, and neither is present here, so adding it would spend latency the budget cannot afford against a problem the reviews show does not exist.
Answers that look right on this material and are not. Each one is a distractor from a different question in the CCDV-F bank for this domain.
Let the full responses accumulate and compact the conversation into a running summary once the window approaches its limit, so the earlier turns are replaced by condensed text.
errorWhy it is wrong: Tempting because compaction is a genuine context-management instrument, but it pays to load 40 kilobytes of unread fields and then pays again to summarise them, treating the symptom rather than the cause.
It removes the need to handle tool errors, because the framework retries any failing tool call until it succeeds and the agent never observes a failed invocation.
errorWhy it is wrong: Frameworks do absorb some retry handling, but an agent has to see a failed test run or a rejected edit, since reacting to that failure is the whole point of the loop.
Manager and subagent hierarchies suit long-running background jobs, so any request that a user waits on should be handled by a single model call instead
errorWhy it is wrong: Tempting because hierarchies do often appear in background work, but the constraint is not whether a user is waiting, and interactive systems can legitimately delegate when the work genuinely divides.