ReAct vs. Plan-Act-Observe-Reflect for a Research Agent

You're building an agent that answers open-ended research questions by searching the web and internal documents, synthesizing across multiple sources, and producing a cited answer. Early testing with a plain ReAct loop shows a specific problem: the agent frequently declares an answer "complete" after finding one source that partially addresses the question, without checking whether it actually answered what was asked or whether contradicting sources exist.

1. Diagnose why a plain ReAct loop is prone to this specific failure.
2. Would adding a plan-act-observe-reflect structure fix it, and specifically which part of that structure addresses this failure?
3. What would make reflection not worth the added LLM calls for this same agent, and how would you scope reflection to avoid that?