Ratchet: A Minimal Hygiene Recipe for Self-Evolving LLM Agents

2026-05-21

AI Summary

This work addresses the performance stagnation in self-evolving large language model (LLM) agents caused by the accumulation of ineffective or redundant skills. It proposes a closed-loop, single-agent system built upon a frozen LLM that autonomously writes, retrieves, organizes, and prunes skills through two core mechanisms: skill elimination and meta-skill-guided refinement. The approach incorporates hygiene mechanisms—including a bounded active skill capacity, a results-driven elimination strategy, and explicit deduplication via meta-skill internalization—to ensure sustained capability growth without performance degradation. Evaluated on MBPP+ hard-100, the method improves pass@1 from 0.258 to a rolling average of 0.584 (peak: 0.658), and achieves a peak improvement of +0.22 on SWE-bench Verified.

Abstract

Self-evolving skill libraries, pioneered by Voyager, let frozen LLM agents accumulate reusable knowledge without weight updates, yet recent evaluation shows that LLM-authored skills deliver $+0.0$pp over no-skill baselines while human-curated ones deliver $+16.2$pp: the bottleneck is not skill authoring but lifecycle management. We introduce \textbf{Ratchet}, a single-agent loop in which a frozen LLM writes, retrieves, curates, and retires its own natural-language skills. Ratchet integrates four candidate hygiene mechanisms: outcome-driven retirement, a bounded active-cap, meta-skill authoring guidance, and pattern canonicalisation. On MBPP+ hard-100 with Claude Opus 4.7, Ratchet lifts held-out pass@1 from a $0.258 \pm 0.047$ baseline to a late-window rolling mean of $0.584$ (peak $0.658 \pm 0.042$) across 100 rounds and 3 seeds, a $+0.328 \pm 0.018$ rolling-mean gain where the no-skill control drifts at $+0.002 \pm 0.005$; the same recipe transfers to an agentic solver on SWE-bench Verified ($+0.22$ peak lift over 20 rounds). Eight ablations (A1--A8) reveal that the minimal working recipe is smaller than our design suggests: retirement and the meta-skill authoring prior are load-bearing, while explicit deduplication (canonicalisation, cover-guard) is subsumed by the meta-skill itself. A non-divergence proposition shows that bounded cap and retirement threshold together prevent expected performance from drifting below the no-skills floor.

Problem

• self-evolving agents
• skill lifecycle management
• LLM-authored skills
• hygiene mechanisms
• performance drift

Innovation

• self-evolving agents
• skill lifecycle management
• LLM hygiene
• meta-skill authoring
• outcome-driven retirement