CCDV-F Practice Questions: Agents and Workflows Domain

Correct answer (B): Managed agent deployment is appropriate when the team wants to reduce setup and maintenance burden and does not need unusual low-level orchestration or custom infrastructure integration. Self-hosting can offer more control, but it adds implementation and operational responsibility that the pilot does not require.

Why the other options are wrong:
- Option A: Self-hosting provides more direct control, but the scenario explicitly says low-level control and custom infrastructure needs are low.
- Option C: A custom framework may be justified for complex orchestration needs, but it conflicts with the pilot goal of minimizing setup and maintenance.
- Option D: Unmanaged scripts may seem simple initially, but they do not provide a coherent agent deployment path and can increase maintenance overhead.

Correct answer (B): A graph or typed-state agent framework is useful when an application benefits from explicit state transitions, structured handoffs, checkpointed retries, and reusable orchestration patterns. The requirements are not just for model reasoning; they are for controlled workflow state and recoverability.

Why the other options are wrong:
- Option A: A free-form loop can support agentic behavior, but it does not naturally provide auditable named states, typed handoffs, or checkpointed retries.
- Option C: A prompt can describe states, but it does not by itself enforce typed transitions, retries, or resumable orchestration.
- Option D: Prior review memory may provide useful context, but it does not implement the required state machine or validation behavior.

Correct answer (A): Hooks provide deterministic application behavior around agent execution. A post-run hook is the right control point when a fixed action must occur after the agent produces the required artifact, rather than relying on the model to remember to perform that step.

Why the other options are wrong:
- Option B: Prompt instructions can guide behavior, but they are not the best mechanism for a required deterministic post-processing action.
- Option C: Memory examples may improve style consistency, but they do not reliably trigger a required application-side update.
- Option D: Subagents address specialization or context isolation, not a missing deterministic post-run action.

Correct answer (A): The Claude Agent SDK is appropriate when the team wants supported abstractions instead of implementing loop, state, and tool-dispatch behavior manually. A custom harness is better reserved for cases needing tight orchestration control or specialized integration behavior.

Why the other options are wrong:
- Option B: A custom harness can be valuable, but the scenario says the team wants supported abstractions and lacks special state-machine requirements.
- Option C: A static prompt does not provide the agent loop and tool-dispatch behavior the team is explicitly seeking.
- Option D: Subagents should be introduced to solve decomposition or context-isolation needs, not before the main task and architecture are defined.

Correct answer (A): The log shows a runaway loop that repeats the same ineffective action. Agent loops should track state, detect repeated observations or failed actions, enforce retry limits, and terminate when no useful progress is being made.

Why the other options are wrong:
- Option B: Persisting full history may preserve evidence, but it does not prevent the current loop from repeating the same failed action indefinitely.
- Option C: Fresh subagents might reduce context contamination, but delegation alone does not create loop termination or repeated-result safeguards.
- Option D: A fixed number of searches limits runtime, but always doing 20 searches is inefficient and does not address progress-based stopping.

Correct answer (D): A supervisor/subagent pattern fits when related subtasks need coordination but benefit from isolated context. The supervisor can assign schema, test, documentation, and rollout subtasks to specialized subagents, then integrate concise summaries or structured results instead of carrying all raw evidence in one monolithic context.

Why the other options are wrong:
- Option A: A larger context can hold more data, but it does not solve context drift from unrelated evidence influencing later reasoning.
- Option B: Independent prompts may isolate work, but without a coordinating parent and handoff summaries, integration and consistency across the migration plan are weaker.
- Option C: Durable memory is for relevant future-use information, not for indiscriminately storing transient migration evidence from the current run.

Correct answer (A): The best design separates deterministic and agentic responsibilities. Fixed, auditable eligibility rules should stay in a deterministic workflow. Ambiguous investigations that require planning, dynamic record selection, observations, and stopping conditions fit an agent loop. This hybrid approach preserves compliance where rules are explicit while allowing autonomy where the path is unknown.

Why the other options are wrong:
- Option B: A single autonomous agent gives unnecessary discretion over fixed regulatory checks that should remain deterministic and auditable.
- Option C: A fully fixed workflow handles rule checks well, but it is too rigid for investigations where the next step depends on observations.
- Option D: A manager-subagent hierarchy adds broad complexity and does not distinguish between fixed rule evaluation and dynamic investigation needs.

Correct answer (A): A custom agent loop is appropriate when the team needs fine-grained control over state, retry behavior, stopping conditions, observability, and handoffs, and is willing to own the implementation burden. The requirements explicitly call for testable execution control and integration with existing telemetry.

Why the other options are wrong:
- Option B: Prompting Claude to describe retries does not provide deterministic, testable control over execution state.
- Option C: A framework can reduce boilerplate, but hiding state transitions conflicts with the stated observability requirement.
- Option D: A deterministic workflow may be testable, but the scenario requires adaptation to observations and varied retry behavior.

Correct answer (A): The log shows repeated identical tool calls despite stable observations, and the loop configuration has no maximum iteration setting or retry limit. A practical Claude agent loop needs explicit termination criteria, iteration bounds, timeout handling, and retry behavior to prevent runaway or repetitive execution.

Why the other options are wrong:
- Option B: Memory may preserve useful durable information, but storing repeated transient observations does not stop the current loop from repeating.
- Option C: A larger payload might provide more detail in some cases, but the evidence already repeats the same result and the missing loop controls are the direct failure mode.
- Option D: Subagents can isolate complex subtasks, but delegating identical lookups would multiply cost and does not address the missing termination controls.

Correct answer (B): A deterministic workflow is best when the sequence of steps, branching rules, and audit requirements are stable. Claude can still be used inside bounded steps, such as extracting fields or drafting messages, while the application enforces approval routing and logging deterministically.

Why the other options are wrong:
- Option A: A free-form agent may seem flexible, but the process does not require dynamic planning and would be harder to audit than fixed branching logic.
- Option C: A manager/subagent design adds coordination overhead without a need for specialized decomposition or context isolation.
- Option D: Memory from prior invoices is not the primary need, and storing approval history in agent memory could weaken auditability.