Multi-agent debate or voting: use it for judgement calls, not for facts
@devsingh05▲ 5 net votes
Running several model calls and aggregating (majority vote, or a judge picking the best) can improve subjective judgements and reduce variance, but it multiplies cost and latency and does not fix missing information. For factual questions, better retrieval or tools beat more agents. Measure on your evaluation set whether the ensemble actually beats a single strong call.