Don't Solve, Just Compare: Tiny Advisors for Runtime Intervention in LLM Agents