Q36 of 60 — CCA Practice Exam

Multi-Agent Research System

Your multi-agent research system generates synthesis reports across 50 source documents nightly. Stakeholders report that roughly 12% of completed reports contain claims that are either unsupported by the cited source or directly contradict findings from another subagent's summary. Logs show each subagent's individual summaries are accurate and well-attributed. The synthesis agent receives all subagent outputs and produces the final report in a single pass. What architectural change would most effectively reduce these cross-citation failures?

A. Increase the synthesis agent's context window allocation so it can hold all 50 source documents alongside the subagent summaries, providing more complete information during final report generation to reduce attribution gaps.

B. Add a stronger system prompt instruction directing the synthesis agent to verify each claim against the provided subagent summaries before including it, turning self-monitoring into an explicit generation requirement.

C. Insert a plan-validate step: require the synthesis agent to produce a structured claims.json mapping each claim to its source subagent, validate that file against subagent outputs for contradictions, then generate the final report.

D. Replace the single synthesis agent with three independent agents that each generate a full report from the same subagent outputs, then apply majority voting across all three outputs to select the final report.