Anthropic CCA-F 問題集

質問 1：
In a multi-agent architecture, what is the best practice for handing off context from a coordinator agent to a specialized subagent?
A. The subagent automatically inherits the coordinator's context window.
B. Grant the subagent access to a shared global state database where the coordinator writes its findings.
C. Share the entire coordinator conversation history to guarantee maximum context.
D. Pass only the explicit context, findings, and parameters specifically relevant to the subagent's assigned subtask.
正解：D
解説: (Topexam メンバーにのみ表示されます)

質問 2：
Which of the following represent architecturally sound strategies for managing context effectively across long documents or multi-agent handoffs? Choose 2 correct answers.
A. Requiring subagents to output structured claim-source mappings to support accurate downstream synthesis.
B. Placing key findings summaries at the beginning of aggregated inputs to mitigate position effects.
C. Replacing specific numeric constraints in long contexts with generalized natural language summaries.
D. Sharing the full coordinator conversation history with every subagent to guarantee maximum context.
正解：A,B
解説: (Topexam メンバーにのみ表示されます)

質問 3：
During the final synthesis phase, the coordinator receives conflicting statistical data from two different subagents (e.g., verified database vs. outdated PDF). What is the correct architectural pattern for handling this conflict?
A. Halt the synthesis process and prompt the user to manually input the correct statistic.
B. Silently average the conflicting numerical values to provide a balanced estimate.
C. Arbitrarily select the value from the most recently updated source to resolve the conflict.
D. Maintain structured claim-source mappings to preserve information provenance (source, confidence, timestamp) for informed resolution.
正解：D
解説: (Topexam メンバーにのみ表示されます)

質問 4：
While working in Claude Code, you press Shift + Tab to enter Plan Mode. Which subagent is typically associated with isolating the verbose discovery phase to prevent context window exhaustion during planning?
A. The Explore subagent
B. The Review subagent
C. The Plan-Executor subagent
D. The Write subagent
正解：A
解説: (Topexam メンバーにのみ表示されます)

質問 5：
Why is configuring the exact same Claude session to review its newly generated pull request code considered an architectural anti- pattern?
A. It exceeds the maximum token limit for the context window during the generation phase.
B. Tools invoked during the first pass are permanently locked and cannot be reused in the review pass.
C. The reviewer retains the reasoning context and assumptions from the generation phase, creating confirmation bias.
D. The Message Batches API does not support multi-turn tool calling within a single request.
正解：C
解説: (Topexam メンバーにのみ表示されます)

質問 6：
In a multi-agent research system, subagent A retrieves revenue data from a verified financial database, while subagent B extracts conflicting revenue data from an outdated PDF. How should the system handle this uncertainty before presenting it to a human reviewer?
A. Average the two values to provide a balanced estimate to the human reviewer.
B. Silently drop the PDF data and return the database value since databases are always correct.
C. Maintain structured claim-source mappings with information provenance (source, confidence, timestamp) so the coordinator or human can make an informed resolution.
D. Restart both subagents using a higher temperature setting to generate a consensus value.
正解：C
解説: (Topexam メンバーにのみ表示されます)