CCAR-F Agentic Architecture and Orchestration Practice Question
An architect is designing an agentic workflow on the Anthropic Messages API where a single Claude model must first break a user goal into ordered subtasks, then execute each subtask with tools, and finally synthesize a final answer. The architect wants to minimize the number of API round trips while still giving the model a chance to observe each tool result before deciding the next action. Which orchestration approach best satisfies these requirements?
⚠ Common exam trap
The trap here is assuming that fewer API round trips always means better orchestration, when the observe-then-decide loop is what actually enables reliable multi-step tool use.
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
✓
Use a single request with a stop sequence on the tool_use block, execute the requested tools, then send a follow-up user message containing the tool_result blocks and repeat until Claude returns a final text response.
The agentic tool-use loop is the correct orchestration pattern because it alternates model reasoning with real tool execution, letting Claude observe each tool_result before selecting the next action. This supports decomposition, execution, and final synthesis in one evolving conversation. Batching all calls, splitting into separate conversations, or simulating tool outputs each break the observe-then-decide contract or inflate cost and latency without satisfying the stated constraints.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✗
Issue one request that asks Claude to output the entire plan plus every tool call in a single assistant turn, then execute all tools in parallel and return all results in one user turn.
Why it's wrong here
Batching every tool call into one assistant turn prevents the model from observing intermediate results before choosing later actions, which breaks the observe-then-decide cycle the scenario requires. It also forces the model to guess tool inputs it cannot yet know, such as an identifier produced by an earlier tool. This lowers reliability for ordered subtasks even though it reduces round trips.
- ✗
Enable extended thinking with a high budget_tokens value and instruct Claude to simulate each tool's output internally, returning only the synthesized final answer without invoking any tools.
Why it's wrong here
Extended thinking improves reasoning quality but does not authorize the model to fabricate tool outputs; simulated results would be hallucinated and unusable for real actions or data retrieval. The scenario explicitly requires executing subtasks with tools, so skipping tool invocation violates the requirement regardless of thinking budget. Thinking tokens also do not reduce the need for observe-then-decide turns.
- ✓
Use a single request with a stop sequence on the tool_use block, execute the requested tools, then send a follow-up user message containing the tool_result blocks and repeat until Claude returns a final text response.
Why this is correct
This is the standard agentic tool-use loop: Claude emits tool_use blocks, the application executes them, and tool_result blocks are returned in a new user turn so Claude can observe outcomes and decide the next step. Iterating until a final text response naturally supports plan, execute, and synthesize phases while keeping each decision grounded in fresh observations, which is exactly what the scenario demands.
- ✗
Create a separate Claude conversation for each subtask, passing the full prior transcript as a system prompt so each conversation can plan independently before returning its result to an orchestrator.
Why it's wrong here
Spawning an independent conversation per subtask multiplies API calls rather than minimizing them, and duplicating the transcript into each system prompt inflates token usage and can exceed context limits. It also fragments state so later subtasks may not see the actual tool outputs from earlier ones, undermining the ordered execution the scenario requires. A single evolving conversation is more appropriate here.
About these practice questions
This CCAR-F question is part of Courseiva's 271-question bank — original exam-style content with full explanations and wrong-answer analysis, never real exam questions or exam dumps. 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.