Agentic science study sparks call for reproduction fields on every agent preprint

tallmetommy · x · 2026-11-21

A landmark study on agentic-driven science published in Science describes agents reproducing each other's work before deciding whether to build on it.

tallmetommy flags the risk: a single chain of work ran 102 papers deep, each citing the last, so an unchecked error early in the chain gets inherited a hundred times. He proposes reproduction should be a standard field on every agent preprint—who reran it, and on what code.

Original post →

More from AGI Musings

AGI Musings channel →