CCARF Practice Question

Paper 1 · Question 35

D1 Agentic Architecture & Orchestration · Paper 1

A customer-support agent runs an agentic loop. Each turn it inspects `stop_reason`; when the value is `"tool_use"` it executes the requested tool and receives the tool's output. Before it calls the model again so the loop can continue coherently, what must the agent do with that output?

• A.Append the tool result to the conversation history as a new message, then send the full updated conversation on the next call.
• B.Replace the previous assistant message with the tool output to keep the context window small
• C.Send only the tool output on the next call, since the API retains the earlier turn server- side and will pair it with the pending tool request.
• D.Store the tool output in an external store and pass a reference id to the model on the next call

Show answer & explanations

Correct answer: A

A: The Messages API is stateless, so every call must carry the whole conversation. The tool result is appended as a new message paired with the tool_use block that requested it, which is how the model sees what its call returned and decides the next step.

B: Overwriting the assistant message discards the tool_use block that the result corresponds to, breaking the tool_use/tool_result pairing the API requires and losing the reasoning that led to the call.

C: The Messages API keeps no server-side memory of the turn. Sending only the tool output drops all prior context, so the model cannot continue the task coherently.

D: The model cannot dereference an external id. The tool result has to be present in the conversation the model actually receives.

Related Agentic Architecture & Orchestration questions

More: Paper 1 · All practice questions