When many agents work the same problem, someone must decide what "the answer" is. Consensus protocols — quorum voting, weighted aggregation, judge hierarchies — turn a swarm's noisy, correlated opinions into a single actionable decision with a quantified confidence.
The foundational result is sobering and useful: majority voting only beats its average member when votes are independent (Condorcet's jury theorem) — and LLM agents fail independence by default. Same model, same prompt, same temperature produces correlated errors that vote together. The engineering remedies are direct: sample with genuine diversity (different models, different personas, different retrieval contexts), decorrelate prompts structurally, and measure agreement structure — a swarm that agrees on everything learns nothing from voting.
Beyond raw voting, production consensus layers add structure. Weighted quorums give historically accurate agents more vote mass — reputation as stake. Supermajority thresholds for irreversible actions: two-thirds agreement to deploy, simple majority for drafts. Deliberate minority protection: a persistent dissenter triggers escalation rather than being outvoted — because in verification settings, the lone agent that noticed the poison pill is the whole point of having a swarm.
Protocol Design Checklist
- Define the vote space: answers must normalize to comparable options before counting — un-normalized free text produces n winners out of n voters.
- Quantify correlation: track agreement matrices across agent pairs; two agents that always agree are one agent with a backup.
- Gate by consequence: consensus thresholds should scale with action irreversibility, not task difficulty.
- Log the dissent: minority reports are training data — for prompts, for routing rules, and for the postmortem when consensus was wrong.
The swarm's real product is not the answer — a single strong model often matches it — but the distribution around the answer: a measurable, auditable confidence that single-agent systems cannot produce at any price.