During interaction, AI agents can accumulate and refine reusable procedural knowledge through self‑evolving skills without changing model parameters. Existing approaches store skills as text and rely on large language model (LLM) reasoning to execute them, so even previously learned steps must be re‑interpreted each time, introducing unavoidable uncertainty. In contrast, program execution is deterministic and independently verifiable. Motivated by this intuition, we introduce SkillCompiler, a framework that incorporates programmatic compilation into skill evolution. SkillCompiler uses an executability‑aware compilation strategy: it decomposes a skill into fine‑grained operations, identifies those with well‑defined executability, and compiles them into validated programs, while retaining LLM reasoning for context‑dependent operations to form hybrid skill skeletons. To avoid redundant synthesis and let executable knowledge accumulate across evolving skills, we add a program asset reuse mechanism that organizes verified programs as reusable assets, retrieves existing implementations when applicable, and only synthesizes and verifies new programs when necessary. Extensive experiments on five benchmarks and six model families show that SkillCompiler consistently outperforms state‑of‑the‑art methods, achieving up to a 32.33% improvement in task performance.
Review