agent-skill-lifecycle

Cross-category topic | 13 sources

Agent Skill Lifecycle #

§1 主题缘起 #

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:

  1. 规模化暴露了退化问题。SkillsBench 实证表明 LLM 自生成技能的提升为 +0.0pp,而人工编写的技能提升为 +16.2pp [2604.15877]。生态系统已达 700K+ 技能 [2605.07358],但未受管理的积累导致 retrieval precision 稀释和 silent injection harm [2605.19576]
    1. 生命周期各阶段孤立演进。对 22 篇核心论文的 1,136 条引用分析显示 memory 社区与 skill 社区的跨社区互引率不到 1% [2604.15877]。创建端(MUSE-Autoskill)、退役端(Library Drift)、路由端(SkillRouter)、优化端(MOCHA)各自独立发展,缺乏统一视角。
      1. 安全与信任问题浮现。42,447 个社区技能中 26.1% 含漏洞 [2602.12430],878 个网络安全 SKILL.md 中仅 2.3% 提供完整的用户可理解性支持 [2605.19362]。快速增长与治理缺失的矛盾使得 lifecycle governance 从"可选"变为"必要"。
      2. 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。

        §2 覆盖的 category 分布 #

        CategoryPaper countRepresentative
        agent12MUSE-Autoskill, SLIM, Library Drift, SkillRouter, SkillsInjector, SkillReducer, SSL, Skill-as-Pseudocode, Agent Skills Architecture, Comprehensive Survey, Experience Compression, Skill Comprehension
        algorithm1MOCHA

        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。

        §3 时间线 #

        时间Entity关键贡献
        2026-02Agent 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]
        2026-03SkillRouter (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]
        2026-03SkillReducer (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]
        2026-04Experience Compression Spectrum (2604.15877)统一 memory/skill/rule 为压缩谱三层级;识别 "missing diagonal" [2604.15877]
        2026-04SSL (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]
        2026-05Comprehensive Survey (2605.07358)122 篇论文 × 四阶段 lifecycle taxonomy;形式化 $S = (M, \mathcal{R}, \mathcal{C})$ [2605.07358]
        2026-05SLIM (2605.10923)Leave-one-skill-out MEC 驱动 non-monotonic skill trajectory (38→21);RL 联合优化 [2605.10923]
        2026-05Library Drift (2605.19576)诊断 library drift 三阶段退化;Ratchet 治理 +0.328 pass@1;Hoeffding evidence floor [2605.19576]
        2026-05Skill 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]
        2026-05MOCHA (2605.19330)Chebyshev scalarization + HVC 探索实现多目标 skill optimization +7.5% correctness [2605.19330]
        2026-05MUSE-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]
        2026-05Skill-as-Pseudocode (2605.27955)Prose→typed pseudocode refactoring +74% 相对胜率同时 −23% tokens [2605.27955]
        2026-05SkillsInjector (2605.29794)Execution-grounded context construction;per-task adaptive budgeting +5.1pp avg [2605.29794]

        §4 Evolution timeline (技术谱系) #

        flowchart TD %% Foundations VOY["Voyager (2023)\nOpen-ended skill accumulation"] --> SR["SkillRL\nRL + monotonic growth"] VOY --> EXPEL["ExpeL\nTwo-speed memory"] %% Architecture & Formalization VOY --> ARCH["Agent Skills Architecture\n(2602.12430)\nSKILL.md + progressive disclosure"] ARCH --> SURV["Comprehensive Survey\n(2605.07358)\nS=(M,R,C) lifecycle taxonomy"] ARCH --> SSL["SSL (2604.24026)\nThree-layer structural graph"] ARCH --> COMP["Skill Comprehension\n(2605.19362)\nFour-anchor audit"] %% Compression & Representation ARCH --> REDUCER["SkillReducer (2603.29919)\nTaxonomy-driven debloating"] REDUCER --> SAP["Skill-as-Pseudocode\n(2605.27955)\nTyped contract + verifier"] %% Retrieval & Selection ARCH --> ROUTER["SkillRouter (2603.22455)\nFull-text listwise reranking"] ROUTER --> INJ["SkillsInjector (2605.29794)\nExecution-grounded selection"] %% Lifecycle Management (the core branch) SR --> SLIM["SLIM (2605.10923)\nMEC + non-monotonic trajectory"] EXPEL --> ECS["Experience Compression\n(2604.15877)\nMissing diagonal"] ECS --> DRIFT["Library Drift (2605.19576)\nRatchet governance"] ECS --> MUSE["MUSE-Autoskill\n(2605.27366)\nFull lifecycle"] SLIM -.-> MUSE DRIFT -.-> MUSE %% Optimization SURV --> MOCHA["MOCHA (2605.19330)\nMulti-objective Pareto search"] %% Cross-links INJ -.->|"$Δ$ as lifecycle signal"| DRIFT SLIM -.->|"MEC portable to frozen-LLM"| DRIFT

        Branch points:

        • Voyager (2023) 分裂为两条路径:monotonic accumulation(SkillRL → SLIM 否定此路径)和 two-speed memory(ExpeL → Experience Compression Spectrum)
        • Agent Skills Architecture (2602.12430) 同时催生了 representation 改进(SSL、Skill-as-Pseudocode)和 retrieval 改进(SkillRouter → SkillsInjector)两条技术线

        Convergence points:

        • MUSE-Autoskill 是 SLIM(value estimation)、Library Drift(governance)、Experience Compression(unified framing)三线汇聚的产物
        • SkillsInjector 吸收了 SkillRouter 的 retrieval + MUSE 的 skill bank 作为上游输入

        §5 技术线交错 #

        Representation 选择约束 Routing 实现 #

        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"。

        Lifecycle Governance 制约 Acquisition 策略 #

        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]

        Algorithm 层面的优化反作用于 Agent Runtime #

        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]

        压缩层级的信息量与 Routing 的区分力 #

        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]

        §6 共识与分歧 #

        共识 #

        1. Skill 积累不等于性能提升。13 篇论文中至少 7 篇明确或隐含地确认:无管控的 skill 积累会退化性能。Full-library injection 导致 pass rate 从 40.2% 崩溃到 24.8% [2605.29794];self-gen skills 在 SkillsBench 上 +0.0pp [2604.15877];SkillRL 的 monotonic accumulation 被 SLIM 直接否定 [2605.10923]
          1. Lifecycle 管理是收益释放的必要条件。全部系统论文(MUSE, SLIM, Library Drift, SkillRouter, SkillsInjector, SkillReducer, Skill-as-Pseudocode)均引入了某种形式的质量/数量治理机制。两个 survey(2605.07358, 2604.15877)也独立得出相同结论 [2605.07358] [2604.15877]
            1. 技能作为可迁移的外部化资产。MUSE 的 cross-agent transfer(MUSE→Hermes 仅差 2pp)[2605.27366]、SkillReducer 的 cross-model retention(0.965 across 5 models)[2603.29919]、Experience Compression 的 compression-as-abstraction 框架 [2604.15877] 共同支持:高质量技能是 agent-agnostic 的知识资产。
            2. 分歧 #

              1. Retirement 的统计门槛。Library Drift 断言 $N_{\min} = 100$(Hoeffding $\epsilon \approx 0.20$)是 load-bearing 的,$N_{\min} = 20$ 导致 bank collapse [2605.19576]。SLIM 使用 $n_{\min} = 20$–30 + EMA smoothing + patience $p = 3$ + 极低 $\tau_{\text{retire}} = 0.001$,达到 87.5% 的最佳成绩 [2605.10923]。矛盾根源在于 SLIM 的三重统计保护的等效强度可能超过 Library Drift 的单一 Hoeffding bound,且 SLIM 的 RL policy 可以补偿误退役 [2605.19576]
                1. Skill 文本应该更长还是更短。SkillReducer 以 $p = 0.002$ 证明 less-is-more [2603.29919];MUSE 的自生成 skill 是人工 skill 的 2.2× 长度,反而节省 20% tokens 和 37% latency [2605.27366]。两者不矛盾——SkillReducer 去除的是 task-irrelevant bloat,MUSE 增加的是 procedural step-by-step specificity——但暗示 "optimal skill length" 取决于内容类型而非绝对 token 数。
                  1. Surface similarity vs execution-grounded selection。SkillRouter 认为 full-text embedding + listwise reranking 即可达 74.0% Hit@1 [2603.22455];SkillsInjector 直接反驳——surface similar skills 可有 $\Delta = -0.20$(如 insurance keyword distractor)[2605.29794]。两者面向不同 scale regime(80K vs ≤82 skills),争议的本质是 scalability-accuracy trade-off [2603.22455]
                    1. RL 训练 vs Frozen model lifecycle。SLIM 将 lifecycle 与 policy gradient 同步优化 [2605.10923];MUSE 和 Library Drift 在 frozen API model 上运行 [2605.27366] [2605.19576]。在 GPT/Claude 等闭源模型为主的部署场景中,RL 方案无法使用 [2605.10923]
                    2. §7 Open challenges (根本性困难) #

                      7.1 跨压缩层级的自适应路由(Missing Diagonal) #

                      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。

                      7.2 大规模长期验证的缺失 #

                      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。

                      7.3 Security × Lifecycle 的统一治理 #

                      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。

                      7.4 Execution-grounded supervision 的 scalability #

                      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。

                      §8 成熟度判断 #

                      整体判断:Research-frontier,accelerating toward production。

                      证据:

                      • 产业信号:MUSE 提及三个 production 系统(SkillMarket、ArkClaw、SkillHub)[2605.27366];技能生态已达 700K+ 规模 [2605.07358];CUA 基准达超人水平 72.6% [2602.12430]
                      • 开源成熟度:13 篇论文中仅 SSL (2604.24026) 和 Skill-as-Pseudocode (2605.27955) 声明开源 [2604.24026] [2605.27955]。核心 lifecycle 系统(MUSE、SLIM、Library Drift)均无公开代码。
                      • 形式化程度:Library Drift 提供 Proposition 1 非发散保证 [2605.19576];SLIM 提供 local lifecycle guarantee [2605.10923];但无 global convergence bound 或 optimal active-set characterization。
                      • Trend direction:从 2026-02 到 2026-05 的四个月内密集出现 13 篇关键论文,论文间互引频繁,形成正反馈循环。领域正从 "scattered solutions" 向 "consolidated lifecycle framework" 过渡。

                      子领域成熟度分解:

                      子方向成熟度证据
                      Skill representation/compressionProduction-readySkillReducer 0.965 cross-model retention [2603.29919]
                      Skill routing at scaleProduction-readySkillRouter 496ms latency, single GPU [2603.22455]
                      Creation-time quality gatingResearch-frontierMUSE unit-test gating 有效但 verifier problem 未解决 [2605.27366]
                      Runtime lifecycle managementResearch-frontierSLIM + Library Drift 提供方向但无统一方案
                      Cross-level adaptive compressionPre-researchMissing diagonal 仅被识别,无系统实现 [2604.15877]
                      Security governanceEarly-stageFramework proposed [2602.12430] 但无公开实现

                      §9 邻接 topic #

                      邻接 Topic边界重叠潜在动态
                      agent-servingSkill routing 的 latency budget 与 inference serving 的 SLA 共享同一 constraintSkill 的 progressive disclosure 本质上是 context-window capacity planning
                      agent-memoryExperience 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-generationSkillRouter 和 SkillsInjector 本质上是 domain-specific RAG pipelineSkill routing 是 RAG 的一个特化实例,但 execution-grounded 信号使其超越标准 RAG
                      multi-objective-optimizationMOCHA 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-security26.1% 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

                      §10 参考 #

                      EntityCategoriesRole in topicKey contribution
                      [2605.27366]agentcreation-first full lifecycleFive-stage closed loop; self-generated skill surpasses human ceiling
                      [2605.10923]agentRL-coupled lifecycle managementLeave-one-skill-out MEC; non-monotonic trajectory; formal local guarantees
                      [2605.19576]agentdrift diagnosis and governanceLibrary drift definition; Ratchet three mechanisms; Hoeffding evidence floor
                      [2604.15877]agentunified theoretical framingExperience compression spectrum; missing diagonal identification
                      [2605.07358]agentcomprehensive lifecycle taxonomy$S = (M, \mathcal{R}, \mathcal{C})$ formalization; 122-paper coverage
                      [2603.22455]agentlarge-scale skill routingFull-text retrieve-and-rerank at 80K scale; listwise reranking
                      [2605.29794]agentexecution-grounded selection$\Delta(t,s)$ supervision; adaptive budgeting; set-aware rendering
                      [2603.29919]agentskill text compressionTaxonomy-driven debloating; less-is-more effect
                      [2605.19330]algorithmmulti-objective skill optimizationChebyshev scalarization; HVC exploration; Pareto front coverage
                      [2604.24026]agentstructural representationThree-layer SSL graph; 98.16% normalization yield
                      [2605.27955]agenttyped pseudocode refactoring4-check deterministic verifier; retrieval-action loop breaking
                      [2602.12430]agentarchitecture and security surveyProgressive disclosure; G1–G4/T1–T4 trust framework
                      [2605.19362]agentuser comprehension auditFour-anchor framework; 92.1%→19.0% steep dropout
                      [2605.27366]agentcross-paper cluster analysisMUSE vs SLIM vs Library Drift positioning
                      [2605.10923]agentcounterfactual vs outcome-based signalsMEC vs $\hat{c}(s)$ comparative analysis
                      [2605.19576]agentevidence floor contradiction analysis$N_{\min}$ calibration debate
                      [2604.15877]agentmissing diagonal critiqueSix-axis critique of compression spectrum framework
                      [2605.07358]agentsurvey gap analysisLifecycle closed-loop as potential artifact
                      [2603.22455]agentrouting paradigm comparisonSurface similarity vs execution utility
                      [2605.29794]agentpipeline positioningContext construction vs retrieval paradigm shift