MOSS: Self-Evolving AI Agents Rewrite Their Own Source Code
A new arXiv paper introduces MOSS, a system enabling autonomous AI agents to rewrite their own source code at the source level, going beyond previous self-evolving agents that only modify text-based artifacts like skill files or prompts. The authors argue that source-level adaptation is more general—Turing-complete, deterministic, and immune to long-context drift—allowing agents to fix structural failures in routing, hook ordering, state invariants, and dispatch that are unreachable from the text layer. MOSS targets production agentic systems, aiming to eliminate recurring failures without human intervention.
Key facts
- MOSS performs self-rewriting at the source level on production agentic systems.
- Previous self-evolving agents only modify text-mutable artifacts (skill files, prompts, memory schemas, workflow graphs).
- Source-level adaptation is Turing-complete and a strict superset of text-mutable scopes.
- It takes effect deterministically rather than through base-model compliance.
- It does not erode under long-context drift.
- Structural failures in routing, hook ordering, state invariants, and dispatch are unreachable from the text layer.
- The paper is published on arXiv with ID 2605.22794.
- The system is designed to learn from user interactions and fix recurring failures autonomously.
Entities
Institutions
- arXiv