Researchers treat AI skills like compiled code—making the same instruction work across different AI models

Computer scientists analyzed 118,000 reusable AI skills and built a system that translates them so they work consistently across different large language models, the way a compiler translates code to run on different processors. In practice, this means an AI agent can use the same skill instruction on eight different models without the skill breaking or behaving unpredictably—and it uses 40% fewer computational tokens to do it.

Right now, AI agents treat skills as raw text that gets pasted into the model's context window. The same skill instruction behaves wildly differently depending on which model you use, which agent framework you run it in, and what environment you're in. SkillRT borrows compiler design—decomposing skills into primitive capabilities, measuring which model-harness pairs support them, then rewriting skills at compile time and optimizing them at runtime. This matters because it solves a real fragmentation problem: as AI agents become production systems with modular, shareable skills, portability and consistency become bottlenecks. If skills can't move reliably across models and platforms, the entire modular-agent ecosystem breaks. This system shows skills can be made portable and efficient—up to 3.2x faster—without rewriting them for each environment.