Multi-step workflows, enforcement & handoffs
Implement multi-step workflows with enforcement and handoff patterns.
This subtopic (1.4) sits in Agentic Architecture & Orchestration (D1) on Anthropic's Claude Certified Architect — Foundations (CCA-F) exam. The bank holds 22 practice questions here — 4 easy, 12 medium, and 6 hard — with 5 free to try, answers and explanations included. 3 of the free questions are below; the rest are in the practice stream.
What the exam tests here
- programmatic enforcement vs prompt-based guidance for workflow ordering
- blocking downstream tools until prerequisite steps complete
- structured handoff protocols — customer details, root cause, recommended actions
- deterministic compliance required for financial operations — prompts have non-zero failure rate
Free practice questions: Multi-step workflows, enforcement & handoffs
Your support agent sometimes calls process_refund using an order ID the customer provided, without first running get_customer and lookup_order to confirm the order belongs to that customer. A handful of refunds have posted to the wrong accounts. What is the most reliable fix?
Show answer & explanation
Correct answer: C. Gate process_refund in the MCP server so it rejects calls lacking a verified customer ID tied to the order.
Enforcement at the tool boundary gives deterministic guarantees that prompting cannot, which matters when refunds have financial consequences. 'Strengthen the process_refund tool description' and 'Add few-shot traces' both rely on probabilistic model compliance. 'Split the workflow into a verification subagent' restructures orchestration but still depends on the refund subagent honoring the handoff contract rather than blocking the tool itself.
The support agent occasionally calls process_refund before lookup_order completes, causing refunds against unverified orders. Which approach most reliably prevents this misordering in production?
Show answer & explanation
Correct answer: A. Enforce programmatically that process_refund cannot be called unless lookup_order has returned a verified order ID.
Programmatic enforcement blocks process_refund until lookup_order returns a verified order ID, guaranteeing sequence regardless of model behavior. 'Update the system prompt...' relies on probabilistic model compliance, insufficient for financial operations. 'Add few-shot examples...' similarly cannot guarantee ordering. 'Log misordering incidents...' is a slow feedback loop that allows errors to continue in the interim.
After escalate_to_human transfers a case, the receiving agent has no customer context and must re-interrogate the customer from scratch. Which change to the escalation handoff most directly solves this problem?
Show answer & explanation
Correct answer: B. Include a structured payload with verified customer ID, issue root cause, and prior steps attempted when invoking escalate_to_human.
A structured handoff payload passed directly to escalate_to_human ensures verified customer details and root cause travel with the transfer deterministically. 'Log the full transcript...' requires a separate manual retrieval step, breaking the handoff flow. 'Prompt the agent to summarize...' relies on the model and the human reading an unstructured message. 'Extend the session timeout...' misunderstands context scope — sessions don't share live state across agents.
2 more free questions on this subtopic in the practice stream, plus 17 in the full bank. Keep practicing →