Overview
Consensus is a count of repeated model outputs, not a human-verified conclusion. The workflow also adds agentic steps, and each step farther from the original instruction can dilute the intended task. A final review should therefore compare the synthesis with the original objective rather than treating majority frequency as an automatic decision.
Outliers still require review. The aggregator should preserve them because a low-frequency answer may expand the search space, but the final decision should distinguish repeated agreement from unusual alternatives.
Additional debate rounds consume time and tokens. They are useful while agents are adding new combinations, objections, or refinements, but less useful once the positions have stabilized.