Preprint proposes rule-based shortcut for routine LLM-agent actions

A preprint describes HaReCAP, an extension to the ReCAP recursive agent framework that compiles frequent successful leaf-level decisions into one-step rules. During use, the system skips an LLM call only when a rule uniquely selects a legal action; otherwise it reverts to ReCAP. Tests using Qwen3.5-27B on Robotouille and ALFWorld reported token-consumption reductions of 14.67%, 17.93%, and 20.08% respectively among tasks solved by both systems. The work characterizes the rules as auditable and able to abstain.