Flat retrieval breaks once a skill library hits the tens-to-hundreds range

Notes on Dynamic Agent Skills: A Lifecycle Survey and Taxonomy of Evolving Skill Libraries (arXiv:2607.10113) — Yubo Li · July 2026

Note published July 25, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

This survey audits 124 papers on agent skill systems published between 2023 and 2026 (2 from 2023, 19 from 2025, 103 from 2026, cutoff May 31, 2026) and builds three shared tools for comparing them. A six-sense taxonomy splits what papers loosely call a "skill" into six structurally different artifacts — executable code, natural-language instructions, SKILL.md-style packages, learned adapters, memory traces, and capability labels — each with its own edit and verification behavior. An eight-stage lifecycle architecture (evidence acquisition, proposal, verification/admission, storage, retrieval/composition, maintenance, distillation, governance) maps the design decisions that recur across systems. A ten-operator vocabulary — Add, Refine, Merge, Split, Prune, Distill, Abstract, Compose, Rewrite, Rerank — gives those changes a common language, instead of treating each update mechanism a paper introduces as a novel method in its own right.

Applying that structure across the corpus turns up seven evidence-graded patterns, not one headline result. Curated, reviewed skills consistently beat unfiltered self-generated ones. Verifier quality is repeatedly decisive for skill-aware reinforcement learning — a weak admission check lets plausible-but-wrong skills into the library, and everything downstream inherits the error. Flat retrieval, meaning search that treats the library as one unstructured pool, degrades once it grows into the tens-to-hundreds-of-skills range; the collapse shows up across several independent systems, not just one. Weaker base models gain disproportionately more from having a skill library than stronger ones do. Narrowly focused libraries often beat broad ones — until the task distribution shifts. And write-time abstraction, structuring a skill well at creation, outperforms leaning on retrieval alone to sort things out later. The survey also catalogs eight named safety surfaces — skill-borne prompt injection, supply-chain poisoning, well-formed-but-harmful skills, credential leakage, among others — and argues that benchmarks should report library trajectories over time instead of single-snapshot performance.