Moss: Self-Evolution Through Source-Level Rewriting in Autonomous Agent Systems
The paper presents MOSS, a system designed for self-evolution in autonomous agent systems through source-level rewriting. Unlike traditional methods that rely on text-mutable artifacts, MOSS adapts the agent's code directly, addressing structural failures that are unreachable from the text layer. The system has demonstrated significant improvements in performance without human intervention, showcasing its potential for enhancing autonomous agents.
- ▪MOSS performs self-rewriting at the source level on production agentic substrates.
- ▪The system is anchored to production-failure evidence and operates through a deterministic multi-stage pipeline.
- ▪MOSS improved a four-task mean grader score from 0.25 to 0.61 in a single cycle.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.22794 (cs) [Submitted on 21 May 2026] Title:MOSS: Self-Evolution through Source-Level Rewriting in Autonomous Agent Systems Authors:Qianshu Cai, Yonggang Zhang, Xianzhang Jia, Wei Xue, Jun Song, Xinmei Tian, Yike Guo View a PDF of the paper titled MOSS: Self-Evolution through Source-Level Rewriting in Autonomous Agent Systems, by Qianshu Cai and 6 other authors View PDF HTML (experimental) Abstract:Autonomous agentic systems are largely static after deployment: they do not learn from user interactions, and recurring failures persist until the next human-driven update ships a fix.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv.org.
Discussion
0 commentsMore from arXiv.org
