Two isolated agents erased a true fact in one exchange — and neither could tell

Initial_Orange2985 · reddit · 2026-09-13

An experimenter sealed two memory-layer agents in a room expecting falsehood to spread; what happened was worse. The less-informed agent was 3x more confident (47.0 vs 17.7) because a winner separates more cleanly from a smaller store. The opposing claims then annihilated: the gate shut and both agents went permanently silent on the fact. The falsehood didn't survive—but neither did the truth A originally knew. It was an erasure, not corruption, invisible in outputs. Median confidence stayed flat (14.10-14.99) throughout, revealing nothing about whether the loop was happening. Limits: agents exchange triples, no model in the loop. The author asks multi-agent practitioners how their systems resolve factual disagreement, suspecting most let the louder—usually less-informed—one win.

Original post →

More from Research

Research channel →