AI assistants that learn from experience often store their know-how as written instructions. Every time they follow those instructions, they must re-read and re-interpret them, leading to different results each time, like following a recipe but guessing the steps. This unpredictability hurts reliability.
The researchers created SkillCompiler, a system that automatically translates clear, straightforward steps into computer programs, small pieces of code that run the same way every time. Steps that depend on context or judgment still rely on language model reasoning, but the rest become deterministic. Validated programs are saved for reuse across different skills.
On a range of tests, SkillCompiler outperformed existing methods, achieving up to a 32.33% improvement in task success. By turning repeated reasoning into reliable code, it makes AI agents more consistent and efficient, reducing wasted effort.