Paper 2 · Question 13

A: While possible, conflicting system prompts would produce consistently wrong outputs across all requests. The more common cause is that the coordinator is not passing the full research output, so the synthesis agent fills gaps with its own training data.

B: Subagents do not share memory. The coordinator must explicitly pass the complete research output to the synthesis agent. If the coordinator only passes a summary, the synthesis agent has no access to the detailed findings and will rely on its training data to fill gaps, potentially contradicting the actual research.

C: A synthesis subagent should not have web search tools; that violates scoped tool access. The real issue is that the full research output is not being passed to it, not that it is searching the web.

D: Model version differences would not cause contradictions with explicitly provided research data. The issue is that the research data is not being fully passed to the synthesis agent, not that the agents have different base knowledge.