Self-evolved skills beat human-curated ones by 17.6 points on SkillsBench

Notes on CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification (arXiv:2604.01687) — Hanrong Zhang, Shichen Fan, Henry Peng Zou, Yankai Chen, Zhenting Wang, Jiayuan Zhou, Chengze Li, Wei-Chieh Huang, Yifei Yao, Kening Zheng, Xue Liu, Xiaoxiao Li, and 1 more · April 2026

Note published July 23, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

AI-assisted notes · reviewed by SkillFed Skill evolution

CoEvoSkills leans on a distinction Anthropic drew when it introduced agent skills: a tool is a single self-contained function, but a skill is a structured bundle of interdependent instructions, scripts, and reference files. Today those bundles are still hand-authored, and it shows — human-curated skills post uneven results on SkillsBench, helpful in most domains but actively degrading performance in others, such as Natural Science. The likely culprit is human-machine cognitive misalignment: a gap between how a person explains a task and how an agent actually needs to execute it. CoEvoSkills replaces manual authoring with two co-evolving LLM sessions: a Skill Generator that rewrites the skill bundle each round, and a Surrogate Verifier — an informationally isolated session that never sees the generator's reasoning or code — which writes and escalates its own test assertions. A held-out ground-truth oracle returns only a pass/fail bit, no test content; that opacity is the mechanism, not a limitation. When the surrogate's own tests pass but the oracle still fails, the blank signal is what forces the verifier to write harder tests, not any direct exposure to what actually went wrong.

On SkillsBench's 87 tasks, CoEvoSkills reaches a 71.1% pass rate on Claude Opus 4.6 with Claude Code — ahead of the no-skill baseline (30.6%), human-curated skills (53.5%), and every self-generation baseline tested, including Anthropic's own skill-creator (34.1%) and a chain-of-thought-guided variant (30.7%). Pull the Surrogate Verifier out of the loop, though, leaving the generator with only the oracle's opaque pass/fail signal, and the pass rate drops to 41.1%. The gain, then, comes from the structured feedback loop itself, not the generation prompt. Skills evolved once on Claude Opus 4.6 transfer to six other models spanning five providers — GPT-5.2, Claude Sonnet 4.5, Claude Haiku 4.5, Qwen3 Coder, DeepSeek V3, and Mistral Large 3 — each picking up 36 to 44 percentage points over its own no-skill baseline. Broken out by domain, self-evolved skills beat human-curated ones in 9 of 11 domains, with the largest swings in Finance and Cybersecurity.

Key numbers

SkillsBench pass rate (Claude Opus 4.6 + Claude Code)	71.1%
Margin over human-curated skills	+17.6pp (53.5% → 71.1%)
Pass rate with Surrogate Verifier removed	41.1% (−30.0pp)
Cross-model transfer gain (6 other LLMs)	+36 to +44pp over each model's no-skill baseline
Average iterations to converge	4.1 verification cycles / 2.4 oracle rounds per task

What this means for skill authors

The finding worth sitting with is the failure mode of human-curated skills, not the win over them. A skill can read as clear, well-organized instruction to the person who wrote it and still cost an agent points relative to having no skill at all — exactly the risk a catalog of agent skills has to screen for. Human review alone won't catch this class of problem, because the failure isn't a bug in the instructions, it's a mismatch between the author's mental model and the executing agent's. That argues for testing skills against actual agent execution, not just reading them, before calling them vetted — the same instinct behind scanning a skill's behavior rather than trusting its documentation.

The cross-model transfer result carries a second, more specific implication: an evolved skill's value looks structural, not memorized. Skills built against one frontier model carried real, if smaller, gains to models across five other labs, meaning a good skill mostly encodes reusable task decomposition rather than one model's quirks. That's the assumption any catalog spanning multiple agent backends has to make to be useful at all, and it holds up here under a controlled comparison. The honest caveat: model-matched evolution still edges out transferred skills by several points, so "portable" isn't the same as "backend-agnostic by default."

The Surrogate Verifier design is worth borrowing from even outside this pipeline: an evaluator deliberately blind to the generator's reasoning and code, escalating its own tests off nothing but a binary pass/fail bit, is a cheap way to keep a verification loop honest without ground-truth labels for every task. Any system that lets an agent author or repair its own skills — a marketplace accepting community submissions included — needs some version of that isolation. Otherwise the reviewer just rubber-stamps whatever the generator already believed.

Skills related to this research

audio-router audio-router intelligently directs audio processing requests to the right specialized skill, whether for playback control, sound analysis, or reactive audio handling. Built for Claude, it streamlines how audio workflows are orchestrated across multiple focused capabilities.★ 8 seedance-recipes Seedance-recipes provides production-ready recipe patterns for video content across genres: product, lifestyle, drama, music video, landscape, commercial, animation, and more. Each recipe preserves core creative constraints while inviting customization of subject, camera, lighting, and sound. Use recipes as proven starting shapes, not rigid templates.★ 5,445 ugc-ads ugc-ads builds short-form product videos optimized for social platforms, combining jump-cut pacing with native lip-synced dialogue delivered as a first-person talking head. The skill auto-detects product category from your URL and structures the narrative across five beats—setup, name reveal, product reveal, twist, and punchline—then renders in your choice of vertical formats.★ 37 10 Music Video This skill transforms your music into precise visual prompts for Seedance 2.0 on Higgsfield, a platform that synchronizes generated video to your uploaded audio track. Learn to reference specific beats, describe genre-matched visuals, and apply proven hooks—from particle explosions on drops to color shifts on harmonic changes—that lock your visuals to every moment of your song.★ 712

Related notes

References

Zhang et al., "CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification," arXiv:2604.01687 (2026)
Li et al., "SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks," arXiv:2602.12670 (2026)
Anthropic, "Agent Skills Overview," 2025
Xia et al., "SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning," arXiv:2602.08234 (2026)
Alzubi et al., "EvoSkill: Automated Skill Discovery for Multi-Agent Systems," arXiv:2603.02766 (2026)