CCAR-F - Agentic Architecture & Orchestration (27% of the exam) - Section 1.4
Implement multi-step workflows with enforcement and handoff patterns.
Programmatic enforcement (hooks, prerequisite gates) versus prompt-based guidance for workflow ordering. Where deterministic compliance is required, such as identity verification before a financial operation, prompt instructions alone carry a non-zero failure rate. Candidates should block downstream tool calls until prerequisites complete, decompose multi-concern requests and investigate each in parallel against shared context, and compile structured handoff summaries for humans who cannot see the transcript.
Human agents receiving escalations from a support resolution agent report that they must re-ask the customer for the order number and for what the agent already attempted. The escalate_to_human tool currently takes a single free-text field, and the raw conversation transcript averages 40 turns and is pasted into the ticket body. Handled-again contacts sit at 22 percent and average handling time on escalated tickets is nine minutes. Which change most effectively addresses the reported symptom?
AAttach the complete raw transcript along with every tool call payload and tool result to the ticket, so the human reviewer has access to the full history of what the agent saw and did before escalating.
BRedefine escalate_to_human with a structured input schema whose required fields are the customer identifier, the order identifier, the issue category, the steps already taken with their outcomes, and the reason automation stopped.check_circle Correct
CMove the agent to a configuration with a substantially larger context window so it can hold the entire conversation in view at the moment it composes the free-text summary it hands to the human reviewer.
DInstruct the agent in the system prompt to write a concise summary covering the customer, the order, the steps already taken and the outstanding question before it calls the escalation tool with its free-text field.
Make a handoff reliable by encoding its required contents as a structured tool schema rather than as a free-text summary instruction. The human reviewer needs specific fields, and a free-text field cannot express that requirement, so omissions stay invisible until the human notices them mid-call. Declaring the escalation tool with required structured inputs moves the contract into the tool definition, where an incomplete call is rejected before a ticket is created. The agent must populate the identifiers and the attempted steps for the escalation to succeed at all, which is what converts a summary habit into a handoff protocol.
Why A is wrong: Tempting because no information is lost and completeness feels safe. It is wrong because the transcript is already in the ticket and the complaint is that it is unusable at 40 turns, so adding payloads increases the reading burden that caused the symptom.
Why B is correct: Correct because a required-field schema turns the handoff into a contract that an incomplete call cannot satisfy, so every escalated ticket arrives carrying the identifiers and the attempt history the human needs.
Why C is wrong: Tempting because summary quality does degrade when relevant turns fall out of context. It is wrong because a 40-turn support conversation fits comfortably already, so the missing fields reflect an unspecified summary contract rather than a capacity limit.
Why D is wrong: Tempting because it names the right content and costs almost nothing to ship. It is wrong because nothing validates the result: a free-text field accepts a summary that silently omits the order number, which is the exact failure being reported.
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 model will call get_customer first, and the SDK will then refuse every other tool call in the session until a customer record has come back successfully from that first invocation.
errorWhy it is wrong: Tempting because it describes the gate the team actually wants, but tool_choice shapes one response and adds no downstream blocking, so that behaviour would need a PreToolUse hook.
Rewrite the CLAUDE.md line as a numbered procedure and repeat the same ordering constraint in the job system prompt, so the required sequence appears twice in the context the model reads on every run.
errorWhy it is wrong: Tempting because restating a constraint more prominently does raise adherence, and CLAUDE.md is the documented place for repository conventions. It is wrong because instructions in context are guidance the model weighs against everything else, so the residual violation rate falls but never reaches the zero the auditor demands.
Attach the full pull request diff to the ticket alongside the job log, so the reviewer has the changed code in the same place as the agent's narrative and does not need to open the repository separately.
errorWhy it is wrong: Tempting because reviewers do need the changed code and consolidating sources is a real convenience. It is wrong because the complaint is about searching through undifferentiated material, and adding a second large artefact increases the volume the reviewer must filter rather than pinpointing the decision.