Tag
MOSS introduces source-level rewriting for self-evolving agents, enabling fixes to structural failures that text-layer evolution cannot reach. It lifts a four-task mean grader score from 0.25 to 0.61 in a single cycle on OpenClaw without human intervention.