Agents
Every 'self-evolving agent' paper this year has mutated text: prompts, skill files, workflow graphs, memory schemas. MOSS from USTC & HKUST …
Every 'self-evolving agent' paper this year has mutated text: prompts, skill files, workflow graphs, memory schemas. MOSS from USTC & HKUST argues this is the wrong layer. The thing that actually brea
Every "self-evolving agent" paper this year has mutated text: prompts, skill files, workflow graphs, memory schemas. MOSS from USTC & HKUST argues this is the wrong layer. The thing that actually breaks agents in production is routing, hook ordering, state invariants, dispatch logic. Those don't live in any text artifact, they live in the agent harness code itself. So text-layer evolution physically cannot reach an entire class of structural failure. MOSS does source-level rewriting on production agentic substrates. Each evolution cycle is anchored to a curated batch of production-failure evidence, code modification is delegated to an external coding-agent CLI, candidates are verified by replaying that batch in ephemeral trial workers, then promoted via consent-gated container swap with health-probe rollback. On OpenClaw the four-task mean grader score lifts from 0.25 to 0.61 in a single cycle without human intervention. Source-level evolution is Turing-complete and a strict superset of any text-mutable scope. The bar for "self-evolving" just moved. MOSS: Self-Evolution through Source-Level Rewriting in Autonomous Agent Systems Paper: https://arxiv.org/abs/2605.22794
Source: Yohei Nakajima (X) | 2026-05-23