Google introduces Procedural Graphs for LLM agents

Unconstrained generation over expanding message histories is hitting hard limits; explicit procedural topologies provide the missing control layer for complex agent execution.

• –Organizes procedural knowledge into editable (procedure, relation, procedure) triplets with localized situational guidance at each decision step
• –Dynamically updates graph topology and edge attributes using an LLM refiner that contrasts successful and failed execution trajectories
• –Retains rejected edits to avoid repeating failure patterns and cyclic errors
• –Consistently outperforms flat-memory and retrieval baselines across multiple agent benchmarks and foundation models