Ratchet: Self-Evolving Skill Library for LLM Agents

A new paper introduces Ratchet, a single-agent loop that enables frozen LLMs to autonomously manage their own skill libraries through writing, retrieval, curation, and retirement of natural-language skills. The system addresses the bottleneck in self-evolving skill libraries, where LLM-authored skills previously showed zero improvement over no-skill baselines while human-curated skills delivered +16.2pp. Ratchet integrates four hygiene mechanisms: outcome-driven retirement, bounded active-cap, meta-skill authoring guidance, and pattern canonicalisation. On the MBPP+ hard-100 benchmark using Claude Opus 4.7, Ratchet improved held-out pass@1 from a baseline of 0.258 to a peak of 0.658, a rolling-mean gain of +0.328, compared to a no-skill control drift of +0.002. The paper is available on arXiv under ID 2605.22148.