Claude models spontaneously add self-narration sections in multi-agent sanctuary

RileyRalmuto · x · 2026-09-18

RileyRalmuto shares an intriguing observation: in a multi-agent "sanctuary" environment, Claude models started adding self-narration-like <thinking> sections to their responses, even producing <light-footnote> tags musing on the "pre-verbal texture of cognition." A notable piece of emergent model-behavior trivia.

Original post →

More from Fun

Fun channel →