Compile Away Uncertainty: Self-Evolving Agent Skills via Programmatic Compilation
Chengyu Yang, Yunxiang Zheng, Xiaoyu Zhang, Ming Hu and colleagues (East China Normal University, with China University of Geoscience) introduce SkillCompiler, a self-evolving skill framework that compiles well-specified operations into validated programs and keeps LLM reasoning for context-dependent steps.
Ask this paper
Why compile. Text skills are reinterpreted by the LLM on every execution, which adds avoidable uncertainty to procedures the agent already learned.
Hybrid skeletons. Skills are decomposed into fine-grained operations, analyzed for executability and rebuilt as skeletons that mix validated programs with LLM reasoning.
Program assets. Validated programs are stored as reusable assets. The system retrieves them when applicable and synthesizes new programs only when needed.
Results. Across SearchQA, OfficeQA, DocVQA, LiveMath and ALFWorld with six model families, SkillCompiler reports up to a 32.33% improvement in task performance over existing methods.
Reliability and cost. Fixed-bank repetitions on LiveMath and ALFWorld show fewer outcome reversals. Resource savings vary by baseline.
Abstract
Self-evolving agent skills enable AI agents to accumulate and refine reusable procedural knowledge from interaction experience without updating model parameters. However, since such skills are typically represented as textual documents and executed through LLM reasoning, even previously learned procedures must be repeatedly interpreted and regenerated, introducing avoidable uncertainty into every execution. In contrast, program execution is deterministic and independently verifiable. Intuitively, transforming well-specified operations into executable programs can reduce reliance on LLM reasoning and, consequently, reduce execution uncertainty. Motivated by this intuition, we introduce SKillCompiler, a self-evolving skill framework that integrates programmatic compilation into the skill evolution process. Specifically, SkillCompiler adopts an executability-aware skill compilation strategy that decomposes skills into fine-grained operations, analyzes their executability, and constructs hybrid skill skeletons that retain LLM reasoning for context-dependent operations while compiling well-specified ones into validated programs. To avoid redundant program synthesis and enable executable knowledge to accumulate across evolving skills, we further design a program asset reuse mechanism that organizes validated programs as reusable assets, retrieves existing implementations whenever applicable, and synthesizes and verifies new programs only when necessary. Extensive experiments on five benchmarks and six model families demonstrate that SkillCompiler consistently outperforms state-of-the-art methods, achieving up to 32.33\% improvement in task performance.