SIGNAL
Multi-agent LLM systems are sold on a built-in second opinion: wire agents together and they can check each other's work. The Niu et al. Reliability-Contagion Feasibility analysis turns that pitch into a tradeoff. In a network where every edge is a chance to pool evidence and a chance to pass a mistake, connectivity is a design lever, not a free good.
The pattern the Niu et al. paper pins is sharper than the headline. In its controlled grok-4.3 experiment, mean first-generation offspring rose from 0.667 to 1.333 to 1.667 as agent degree climbed from 2 to 4 to 5, even as the share of exposed neighbors taking the error held flat at 0.333. Same fraction, more neighbors reached. Adding edges multiplies a single mistake's reach rather than simply adding capacity.
The strongest counterargument lives inside the Niu et al. paper. Hold the sender's budget fixed instead of the per-edge budget, and the early-propagation threshold stops depending on network density. The convention you choose, per edge or per sender, is the lever. Most reader-facing coverage will read the work as a warning about agent graphs; the mechanism says it is a question about which budget you hold fixed, and a tractable question at that.
For buyers and builders, the stake is direct. Under a fixed per-edge exposure budget, reliability and error control in LLM multi-agent networks can pull in opposite directions, and the math says their intersection can be empty. The edges that check the work are the same edges that carry the mistake.