Procedural Graphs Paper Turns LLM Agent Playbooks Into Self-Evolving Graphs of Procedure Triplets

Yuxing Lu and coauthors (including Google Cloud's Sercan Arık) propose Procedural Graphs, encoding LLM agent 'what-to-do' knowledge as (procedure, relation, procedure) triplets. A guidance model reads the local subgraph and emits step-level situational hints, while a refiner LLM inspects failed vs. successful trajectories and rewrites the graph's structure — letting agents bootstrap from minimal templates or repair broken ones. The 36-page paper reports consistent gains over memory-based baselines across multiple datasets.