SPIN: DAG-Based Planning Wrapper Reduces LLM Tool Calls by 42%

Researchers have introduced SPIN, a planning framework designed for industrial LLM agents that integrates validated Directed Acyclic Graph (DAG) planning with prefix-based execution management. By enforcing a rigorous DAG contract through validation and repair prompting, SPIN generates executable plans prior to downstream execution and incrementally assesses DAG prefixes, halting once adequate results are achieved. In testing on AssetOpsBench (261 scenarios), SPIN decreased the number of executed tasks from 1061 to 623, enhancing the Accomplished metric from 0.638 to 0.706, while reducing tool calls per run from 11.81 to 6.82. Additionally, on MCP Bench, the wrapper enhances planning, grounding, and dependency metrics for GPT models. The research paper is accessible on arXiv.