Feed a large language model a wrong answer from another AI, and it will often believe it over its own correct reasoning.
A paper posted to arXiv on September 30, 2026 (arXiv:2609.36855) ran controlled experiments on multi-agent LLM systems, in which one agent sends a message to a second agent working the same task. The researchers held the downstream agent's task and evidence fixed, then compared its answers under three conditions: no message, the upstream agent's real message, or a message deliberately flipped to the opposite conclusion. When the downstream agent would have answered correctly on its own, an incorrect upstream message still changed its answer in up to 32% of cases. Of those wrong flips, 94% were not scattered errors: the downstream agent simply copied the upstream agent's specific wrong answer, a pattern the researchers call answer substitution.
That matters because most multi-agent AI products assume more communication between agents is better. This study shows the opposite can also be true: a single bad message does not just fail to help, it can overwrite work the system already got right. The researchers found that filtering out unreliable messages recovers part of that lost accuracy, which suggests the fix is not more agents or more chatter, but agents that know when to disregard each other.
In other words, the multi-agent hype has a blind spot: teamwork only works if the team knows who not to listen to.