A new technique lets AI planning agents skip redundant reasoning steps, cutting their token bills without sacrificing accuracy.
Researchers describe HaReCAP, an extension to a recursive planning framework called ReCAP that helps large language model agents complete long, multi-step tasks like cooking or fetching objects in a simulated house. The problem: even after ReCAP breaks a goal into subtasks, the agent still calls the LLM at every single leaf-level step to translate a subtask into a valid action, a repetitive process the authors call last-mile grounding redundancy. HaReCAP fixes this by mining common leaf decisions from past successful runs and turning them into simple, auditable rules. At runtime, the agent uses a rule instead of a fresh LLM call whenever one rule can uniquely pick the legal action, otherwise it falls back to the original ReCAP process.
Token consumption is the real cost driver for agentic AI systems that loop through many LLM calls per task, and most efficiency work targets the planning layer rather than this repetitive last-mile step. On Robotouille and ALFWorld benchmarks with Qwen3.5-27B, HaReCAP cut token use by 14.67% to 20.08% on tasks both systems solved, without changing the underlying control flow. That is a meaningful dent in inference cost for anyone running embodied or long-horizon agents at scale.
It is a narrow fix for a narrow problem, tested on two research benchmarks rather than production systems, but the underlying idea, cache the boring decisions and save the model for the hard ones, is the kind of pragmatic optimization agentic AI will need a lot more of.