The paper introduces Procedural Graphs, a framework that organizes procedural knowledge as structured (procedure, relation, procedure) triplets to guide LLM agent decision-making at each step. The system localizes an agent’s position within the graph and offers situational guidance that influences, without dictating, the next action, addressing agents losing track of objectives during extended reasoning. The graph self-improves through an LLM refiner that compares failed and successful trajectories and edits the topology automatically, showing consistent gains over memory-based baselines across multiple datasets and models.
