Moss: Self-Evolution Through Source-Level Rewriting in Autonomous Agent Systems
12 hours ago
- #autonomous agents
- #source rewriting
- #self-evolution
- MOSS is a system enabling autonomous agents to self-evolve through source-level rewriting of their own code.
- It addresses the limitation of static post-deployment agents by allowing modifications at the code level, not just text artifacts like prompts or memory schemas.
- The self-evolution process is anchored to automatically curated batches of production-failure evidence and uses a deterministic multi-stage pipeline.
- Code modifications are delegated to an external coding-agent CLI, with MOSS managing stage ordering and verification.
- Candidate code changes are verified by replaying failure evidence in trial workers, then promoted via user-consent-gated container swaps with health-probe-gated rollback.
- On the OpenClaw benchmark, MOSS improved a four-task mean grader score from 0.25 to 0.61 in one cycle without human intervention.