SkillSmith: Compiling Agent Skills into Boundary-Guided Runtime Interfaces

SkillSmith：将智能体技能编译为边界引导的运行时接口

Recently, skills have been widely adopted in large language model (LLM)-based agent systems across various domains. In existing frameworks, skills are typically injected into the agent reasoning loop as contextual guidance once matched to a runtime task, enabling specialized task-solving capabilities. 近年来，技能已被广泛应用于各领域基于大语言模型（LLM）的智能体系统中。在现有的框架中，技能通常在匹配到运行时任务后，作为上下文引导注入到智能体的推理循环中，从而赋予其专门的任务解决能力。

We find that this execution paradigm introduces two major sources of redundancy: irrelevant context injection and repeated skill-specific reasoning and planning. To this end, we propose SkillSmith, a boundary-first compiler-runtime framework that compiles skill packages offline into minimal executable interfaces. 我们发现，这种执行范式引入了两个主要的冗余来源：无关的上下文注入以及重复的技能特定推理与规划。为此，我们提出了 SkillSmith，这是一个以边界为先的编译器-运行时框架，它能在离线状态下将技能包编译为最小化的可执行接口。

By extracting fine-grained operational boundaries from skills, SkillSmith enables agents to dynamically access and execute only the relevant components at runtime, thereby minimizing unnecessary context injection and redundant reasoning overhead. 通过从技能中提取细粒度的操作边界，SkillSmith 使智能体能够在运行时动态地仅访问和执行相关组件，从而最大限度地减少不必要的上下文注入和冗余的推理开销。

In the evaluation on SkillsBench benchmark, SkillSmith reduces solve-stage token usage by 57.44%, thinking iterations by 42.99%, solve time by 50.57% (2.02x faster), and token-proportional monetary cost by 57.44% compared with using raw-skills. 在 SkillsBench 基准测试的评估中，与使用原始技能相比，SkillSmith 将求解阶段的 Token 使用量减少了 57.44%，思维迭代次数减少了 42.99%，求解时间缩短了 50.57%（速度提升 2.02 倍），且与 Token 成比例的货币成本降低了 57.44%。

Moreover, compiled artifacts produced by a stronger model can be reused by a smaller or more efficient runtime model, improving task accuracy in cases where raw skill interpretation fails. The source code and data are available at this https URL. 此外，由更强模型生成的编译产物可以被更小或更高效的运行时模型复用，从而在原始技能解释失败的情况下提高任务准确率。源代码和数据可在该链接获取。