CCAR-F Agentic Architecture and Orchestration Practice Question
A claims-processing agent must extract structured fields from scanned documents, then decide whether to auto-approve, request more evidence, or escalate to a human. The architect wants deterministic, auditable control over which branch executes rather than letting Claude narrate the decision. Which TWO design choices best achieve that? (Choose two.)
⚠ Common exam trap
The trap here is treating a reliably phrased sentence from the model as equivalent to a validated structured decision, when only typed outputs can be rejected before an irreversible action fires.
Answer choices
Why each option matters
Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.
Correct answer & explanation
✓
Have Claude return a JSON object with a decision field and validate it against a schema before acting on it.
Deterministic, auditable branching comes from making the decision an explicit structured signal the orchestrator can inspect and validate before acting. Either a small set of named tools or a schema-validated decision field turns the model's choice into a typed, loggable value, whereas text conventions, temperature tuning, and unmediated API access all leave the branch implicit or uncontrolled.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✗
Allow Claude to call any of the organization's internal APIs directly so it can decide the branch at execution time without orchestrator involvement.
Why it's wrong here
Granting broad direct API access removes the orchestrator's ability to gate and log the branch decision, and it widens the blast radius if the model reasons incorrectly. It also makes the path taken dependent on external side effects rather than a recorded decision. Auditable routing requires the branch to be an explicit, inspectable signal before any consequential action executes.
- ✗
Instruct Claude in the system prompt to begin every response with the word APPROVE, EVIDENCE, or ESCALATE, then parse that prefix in application code.
Why it's wrong here
Parsing a required text prefix couples routing to fragile string handling and gives no schema enforcement on arguments. The model can emit the prefix inconsistently, wrap it in extra text, or choose a label that does not match the actual content, producing silent misroutes. Structured tool calls provide the same semantic decision with typed, validated parameters rather than a textual convention.
- ✓
Have Claude return a JSON object with a decision field and validate it against a schema before acting on it.
Why this is correct
A schema-validated decision field gives the orchestrator a single typed value to switch on and rejects malformed outputs before any action fires. Validation turns an otherwise soft model output into a contract, so unexpected or missing fields fail loudly instead of triggering an unintended branch. Combined with a defined enum of decisions, this yields deterministic, auditable routing.
- ✗
Set temperature to its maximum so the agent explores all possible decision branches before committing to one.
Why it's wrong here
Higher temperature increases sampling randomness and makes branch selection less reproducible, which is the opposite of deterministic, auditable routing. For classification-style decisions the architect wants low variance so identical inputs map to identical branches. Sampling temperature governs token randomness, not the set of available actions, so it cannot create or constrain decision paths.
- ✓
Define discrete tools such as approve_claim, request_evidence, and escalate_to_human, and use the tool the model selects as the routing signal.
Why this is correct
Each tool name is a finite, machine-readable branch, so the orchestrator inspects which tool_use block Claude returned and routes accordingly. This makes the decision auditable because the branch is recorded as a tool call with typed arguments rather than buried in prose, and it constrains the model to a known set of outcomes instead of free-form narration.
About these practice questions
One of 271 original CCAR-F practice questions on Courseiva, each with a full explanation and wrong-answer analysis — not exam dumps or protected exam content. Learn why practice questions differ from exam dumps →
JA
Written and reviewed by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
Last reviewed September 2026 · checked against the official Anthropic exam blueprint
This CCAR-F practice question is part of Courseiva's free Anthropic certification practice question bank. Courseiva provides original exam-style practice questions with explanations, topic-based practice, mock exams, readiness tracking, and study analytics to help learners prepare for the CCAR-F exam.