Moss: Self-Evolution Through Source-Level Rewriting in Autonomous Agent Systems (arxiv.org)

Researchers have introduced MOSS, a groundbreaking system that enables self-evolution in autonomous agent systems through source-level rewriting. Unlike traditional systems that rely on text-mutable artifacts for updates, MOSS addresses the limitations of static agents by allowing adaptations at the source code level. This approach is significant for the AI/ML community because it tackles a fundamental challenge of structural failures that are unreachable using conventional methods. By being Turing-complete and deterministic, MOSS ensures that adaptations effectively improve agent performance without the risk of degradation over time. MOSS functions through a multi-stage pipeline that utilizes production-failure evidence to guide source-level adaptations. It uses a pluggable external coding agent for modifications while maintaining control over stage order and outcomes. Verification is achieved by testing changes against past failures using ephemeral trial workers, ensuring a robust process for rollouts. Remarkably, in tests with OpenClaw, MOSS improved the system’s mean grader score from 0.25 to 0.61 in a single cycle, demonstrating its potential to enhance autonomous agents autonomously, reducing dependence on human intervention and accelerating deployment cycles in real-world applications.