A team of researchers from Google, Georgia Institute of Technology, and Peking University have put forward the concept of Procedural Graphs (PG) as a solution to the challenges of coordination and experience reuse in intelligent agents. These challenges specifically arise in the contexts of tool invocation, skill utilization, and memory preservation during task execution. PG effectively links components like tools, skills, and memories into a cohesive network. It employs 'process-relation-process' triplets to precisely articulate the conditions for utilizing capabilities, the methods for connecting them, and the pitfalls to steer clear of. This approach enables the retention of procedural memory without the necessity of retraining model weights when modifications are made.

During the execution of tasks, the intelligent agent receives action guidance by pinpointing the current step, extracting pertinent local subgraphs, and formulating situational instructions. In the offline phase, the system refines the procedural graph through a continuous cycle of execution, analysis, modification, and validation, thereby preserving experiences that have been verified. Experiments have demonstrated that PG significantly boosts the performance of intelligent agents in various scenarios, including tool invocation, specialized tasks, and long-term decision-making. While guidance from local subgraphs enhances success rates and cuts down on Token consumption compared to relying on the full graph, it also introduces certain additional overheads, with its benefits being somewhat limited in certain tasks.