Multi-Agent vs. Single Agent

Value criterion: Does the interaction process introduce new information that a single agent would not have obtained during generation? Multiple agents discussing a single text (debate) – with equal budget, not better than one (data processing inequality: sequential passing only loses information). Reviewer with external feedback (test results, rendering screenshots, tool calls) – significant improvement: RLEF, WebGen-Agent (26.4% → 51.9%). Argument boundaries: does not extend to independent multiple sampling and the asymmetry of “easier to check than to generate”. Cost: multi-agent system is ~15x more expensive than dialogue; tokens explain ~80% of the performance difference.

Related: [Proposer-Reviewer], [Interaction Topologies], [Failure Modes of Multi-Agent Systems]