The verifiability litmus test for agent design

The most critical factor in designing an effective AI agent is whether its output can be definitively verified. For tasks with clear, checkable answers like agentic search, the optimal strategy is to invest in verification and selection to pick the single best result from many attempts. However, when an answer's quality is subjective and cannot be easily checked, such as creating a comprehensive research report, the approach must change. In these cases, the best strategy is to generate diverse outputs and then aggregate or merge them, as no single answer can be trusted as complete. This "verifiability litmus test" allows for more efficient agent optimization before resorting to simply scaling compute power.