Agent skill 在 2024–2025 年间从"有就好"(static injection)进入"怎么管"(dynamic lifecycle management)阶段。这一转折有三个触发事件:
本 topic 跨越 agent(系统级 lifecycle 工程)和 algorithm(多目标优化、RL 耦合训练)两个 category,统一审视技能从诞生到退役的完整管线。
| Category | Paper count | Representative |
|---|---|---|
| agent | 12 | MUSE-Autoskill, SLIM, Library Drift, SkillRouter, SkillsInjector, SkillReducer, SSL, Skill-as-Pseudocode, Agent Skills Architecture, Comprehensive Survey, Experience Compression, Skill Comprehension |
| algorithm | 1 | MOCHA |
技能生命周期研究高度集中在 agent category(runtime 行为治理),仅 MOCHA 将 skill text 的优化形式化为 algorithm category 下的多目标搜索问题。但多个 agent-category 论文使用了 algorithm 层面的工具(SLIM 的 GRPO、SkillsInjector 的 preference learning、SkillRouter 的 contrastive learning),暗示两个 category 的边界在此 topic 中模糊化。
| 时间 | Entity | 关键贡献 |
|---|---|---|
| 2026-02 | Agent Skills Architecture (2602.12430) | 首篇系统综述 SKILL.md 范式;三层渐进加载架构 + G1–G4/T1–T4 安全治理框架 [2602.12430] |
| 2026-03 | SkillRouter (2603.22455) | ~80K 池中 full-text retrieve-and-rerank 达 74.0% Hit@1;证明 metadata-only 下降 31–44pp [2603.22455] |
| 2026-03 | SkillReducer (2603.29919) | 首个 build-time skill debloating 框架;48% description + 39% body compression 同时提升 2.8% 质量 [2603.29919] |
| 2026-04 | Experience Compression Spectrum (2604.15877) | 统一 memory/skill/rule 为压缩谱三层级;识别 "missing diagonal" [2604.15877] |
| 2026-04 | SSL (2604.24026) | 三层结构化图表示(调度/场景/逻辑)使 Skill Discovery MRR@50 +0.080 [2604.24026] |
| 2026-05 | Comprehensive Survey (2605.07358) | 122 篇论文 × 四阶段 lifecycle taxonomy;形式化 $S = (M, \mathcal{R}, \mathcal{C})$ [2605.07358] |
| 2026-05 | SLIM (2605.10923) | Leave-one-skill-out MEC 驱动 non-monotonic skill trajectory (38→21);RL 联合优化 [2605.10923] |
| 2026-05 | Library Drift (2605.19576) | 诊断 library drift 三阶段退化;Ratchet 治理 +0.328 pass@1;Hoeffding evidence floor [2605.19576] |
| 2026-05 | Skill Comprehension (2605.19362) | 首次从 usable-security 角度审计 SKILL.md;四锚框架揭示 92.1%→19.0% 陡降 [2605.19362] |
| 2026-05 | MOCHA (2605.19330) | Chebyshev scalarization + HVC 探索实现多目标 skill optimization +7.5% correctness [2605.19330] |
| 2026-05 | MUSE-Autoskill (2605.27366) | 五阶段闭环(创建→记忆→管理→评估→迭代);自生成 skill 超越人工上限 87.94% [2605.27366] |
| 2026-05 | Skill-as-Pseudocode (2605.27955) | Prose→typed pseudocode refactoring +74% 相对胜率同时 −23% tokens [2605.27955] |
| 2026-05 | SkillsInjector (2605.29794) | Execution-grounded context construction;per-task adaptive budgeting +5.1pp avg [2605.29794] |
Branch points:
Convergence points:
SSL 的三层结构化表示使 Skill Discovery MRR@50 从 0.649 提升至 0.729 [2604.24026],而 SkillRouter 证明仅 metadata(name + description)在 80K 池中导致 31–44pp Hit@1 下降 [2603.22455]。两者共同揭示:技能的表示格式直接决定了 routing 的性能上限。Skill-as-Pseudocode 进一步证明从 prose 到 typed pseudocode 的转换同时改善了 agent 的 post-retrieval action 成功率(20.5% → 29.3%)[2605.27955]——表示格式不仅影响"能否找到",还影响"找到后能否用对"。
Library Drift 的 meta-skill authoring prior 是单一最有价值组件(贡献 43% 增益)[2605.19576],表明 skill 创建端的质量约束比退役端的清理更重要。MUSE-Autoskill 的 unit-test gating(入口端)和 Library Drift 的 outcome-driven retirement(出口端)从两端解决同一问题,但两者从未被组合验证 [2605.19576]。
MOCHA 发现 single-objective baseline 在 4/6 任务上 1000 rollout 零进展 [2605.19330]——选择策略而非 mutation 质量是瓶颈。这一发现直接适用于 agent runtime 中的 skill selection:SkillsInjector 的 single-scalar 目标(task success)可能面临类似困境,当 correctness 与 compliance(如 token budget、interaction cost)冲突时需要多目标框架 [2605.29794]。
SkillReducer 发现 body 中仅 38.5% 是 actionable core rules,压缩 39% 反而提升 2.8% 质量 [2603.29919]。SkillRouter 需要 full body text 作为 routing signal [2603.22455]。这表面矛盾实际反映了不同阶段的不同需求:routing 需要 full text 用于 embedding discrimination,execution 需要 compressed text 避免 attention dilution [2603.22455]。
将经验压缩到哪个层级(episodic memory / procedural skill / declarative rule)目前缺少自动决策机制。所有 20+ 已有系统固定在单一压缩层级操作 [2604.15877]。需要 agent + algorithm 两个 category 协同:algorithm 层面设计跨层级的 reward metric($L_1$ 的 retrieval relevance、$L_2$ 的 task success、$L_3$ 的 cross-task generalizability 不可直接比较),agent 层面实现 idle-time consolidation 和 failure-triggered demotion [2604.15877]。无当前方法能完整解决;预计困难度中等,2–3 年内可能出现 partial solutions。
所有系统的实验规模有限:Library Drift 100 轮 × 100 task, SLIM 120–180 RL steps, MUSE 51 task, SkillRouter 75 queries [2605.19576]。真实 agent 部署(coding copilot、customer-facing 助手)运行数月并累积数千技能。在此规模下:meta-skill 是否仍能 subsume explicit dedup?active-cap 是否需要自适应调节?Non-monotonic trajectory 是否收敛到稳态还是持续振荡?这些问题需要 production telemetry 而非 benchmark 实验。
26.1% 社区技能含漏洞 [2602.12430],安全治理(G1–G4 门控 + T1–T4 信任等级)和 lifecycle 管理(creation → evaluation → retirement)目前完全分离。安全审计关注 adversarial behavior,lifecycle 管理关注 performance contribution——两者各有盲区。一个通过所有安全审计的 skill 仍可能因为 stale 内容导致 performance regression,一个 high-MEC skill 仍可能含有 data exfiltration 漏洞 [2602.12430]。统一治理需要同时理解 NL 指令的语义安全性和运行时行为的贡献价值。
SkillsInjector 的 $\Delta(t,s)$ 信号需要 $N_{\text{skills}} \times M_{\text{tasks}} \times K_{\text{seeds}}$ 次 agent rollout——仅 tau2-airline 一个 domain 就需 20,500 次 multi-turn interaction [2605.29794]。在 700K+ 技能生态中,combinatorial data-generation cost 使纯 rollout-derived supervision 不可行。需要 cheap proxies(如 per-skill memory 积累的 success/fail 计数 [2605.29794])或 transfer learning across domains,但两者均未被验证。
整体判断:Research-frontier,accelerating toward production。
证据:
子领域成熟度分解:
| 子方向 | 成熟度 | 证据 |
|---|---|---|
| Skill representation/compression | Production-ready | SkillReducer 0.965 cross-model retention [2603.29919] |
| Skill routing at scale | Production-ready | SkillRouter 496ms latency, single GPU [2603.22455] |
| Creation-time quality gating | Research-frontier | MUSE unit-test gating 有效但 verifier problem 未解决 [2605.27366] |
| Runtime lifecycle management | Research-frontier | SLIM + Library Drift 提供方向但无统一方案 |
| Cross-level adaptive compression | Pre-research | Missing diagonal 仅被识别,无系统实现 [2604.15877] |
| Security governance | Early-stage | Framework proposed [2602.12430] 但无公开实现 |
| 邻接 Topic | 边界重叠 | 潜在动态 |
|---|---|---|
| agent-serving | Skill routing 的 latency budget 与 inference serving 的 SLA 共享同一 constraint | Skill 的 progressive disclosure 本质上是 context-window capacity planning |
| agent-memory | Experience Compression Spectrum 明确统一 memory ($L_1$) 与 skill ($L_2$) [2604.15877];MemSkill 跨越两个 topic | 未来可能合并为 "agent-experience-management" |
| retrieval-augmented-generation | SkillRouter 和 SkillsInjector 本质上是 domain-specific RAG pipeline | Skill routing 是 RAG 的一个特化实例,但 execution-grounded 信号使其超越标准 RAG |
| multi-objective-optimization | MOCHA 的 Chebyshev scalarization 和 SkillsInjector 的 dual-loss planner | 如果 skill optimization 普遍采用多目标框架,可能分裂出独立 topic |
| agent-security | 26.1% 漏洞率和四锚 comprehension framework 与 skill lifecycle 强耦合 | 当前作为 lifecycle 的横切关注点;若产出更多 security-specific 研究可能独立 |
| Entity | Categories | Role in topic | Key contribution |
|---|---|---|---|
| [2605.27366] | agent | creation-first full lifecycle | Five-stage closed loop; self-generated skill surpasses human ceiling |
| [2605.10923] | agent | RL-coupled lifecycle management | Leave-one-skill-out MEC; non-monotonic trajectory; formal local guarantees |
| [2605.19576] | agent | drift diagnosis and governance | Library drift definition; Ratchet three mechanisms; Hoeffding evidence floor |
| [2604.15877] | agent | unified theoretical framing | Experience compression spectrum; missing diagonal identification |
| [2605.07358] | agent | comprehensive lifecycle taxonomy | $S = (M, \mathcal{R}, \mathcal{C})$ formalization; 122-paper coverage |
| [2603.22455] | agent | large-scale skill routing | Full-text retrieve-and-rerank at 80K scale; listwise reranking |
| [2605.29794] | agent | execution-grounded selection | $\Delta(t,s)$ supervision; adaptive budgeting; set-aware rendering |
| [2603.29919] | agent | skill text compression | Taxonomy-driven debloating; less-is-more effect |
| [2605.19330] | algorithm | multi-objective skill optimization | Chebyshev scalarization; HVC exploration; Pareto front coverage |
| [2604.24026] | agent | structural representation | Three-layer SSL graph; 98.16% normalization yield |
| [2605.27955] | agent | typed pseudocode refactoring | 4-check deterministic verifier; retrieval-action loop breaking |
| [2602.12430] | agent | architecture and security survey | Progressive disclosure; G1–G4/T1–T4 trust framework |
| [2605.19362] | agent | user comprehension audit | Four-anchor framework; 92.1%→19.0% steep dropout |
| [2605.27366] | agent | cross-paper cluster analysis | MUSE vs SLIM vs Library Drift positioning |
| [2605.10923] | agent | counterfactual vs outcome-based signals | MEC vs $\hat{c}(s)$ comparative analysis |
| [2605.19576] | agent | evidence floor contradiction analysis | $N_{\min}$ calibration debate |
| [2604.15877] | agent | missing diagonal critique | Six-axis critique of compression spectrum framework |
| [2605.07358] | agent | survey gap analysis | Lifecycle closed-loop as potential artifact |
| [2603.22455] | agent | routing paradigm comparison | Surface similarity vs execution utility |
| [2605.29794] | agent | pipeline positioning | Context construction vs retrieval paradigm shift |