Compile, Then Page: Executable SOP Programs and a Capability-Gated Runtime for Procedural LLM Agents

arXiv cs.AI · 6 hours ago

Researchers developed a system that compiles standard operating procedures into executable pseudo-code and runs them with a program-guided stack machine to control LLM agent behavior. Testing across six models showed compiled SOPs improved performance by up to 16 points on stronger models while harming weaker ones, with the Bank domain achieving 92.8 performance and 100% refusal correctness. The approach demonstrates that runtime guidance effectiveness depends on model capability, with recommendations to compile procedures first and enable advanced features only after confirming model discipline.